ITNG 2022 19th International Conference on Information Technology-New Generations (Advances in Intelligent Systems and Computing, 1421) 3030976513, 9783030976514
This volume represents the 19th International Conference on Information Technology - New Generations (ITNG), 2022. ITNG
111 15 18MB
English Pages 420 [391] Year 2022
Contents
Chair's Message
ITNG 2022 Reviewers
Part I Software Engineering
1 Cross-Platform Blended Modelling with JetBrains MPS and Eclipse Modeling Framework
1.1 Introduction
1.2 Motivation and Contribution
1.3 Related Work
1.4 Bridging MPS and EMF
1.4.1 EcoreLanguage: Implementing Ecore in MPS
1.4.2 Automatic Export from MPS to EMF
1.5 Evaluation
1.6 Conclusions and Future Work
References
2 A Conceptual Framework for Software Modeling of Automation Systems
2.1 Introduction
2.2 Technologies Comprising the Framework
2.2.1 Time-Sensitive Networking
2.2.2 OPC UA
2.2.3 Languages and Models for Software Development
2.3 Proposed Framework
2.3.1 The Conceptual Communication Infrastructure
2.3.2 Holistic Modeling Approach
2.4 Conclusion and Future Work
References
3 Virtual Reality Multiplayer Interaction and Medical Patient Handoff Training and Assessment
3.1 Introduction
3.2 Background and Related Work
3.2.1 Medical Training Simulators
3.2.2 VR in Medicine
3.2.3 Multiplayer VR Simulators
3.3 Implementation
3.3.1 Process of Execution
3.3.2 Eye Tracking
3.3.3 Recording
3.4 Room Creation
3.4.1 Photogrammetry
3.4.2 Room Objects
3.5 Interaction
3.5.1 Avatars
3.5.2 Voice and Audio
3.5.3 Clipboard
3.6 Framework
3.6.1 Interactables
3.6.2 Voice
3.7 Conclusions
3.8 Future Work
3.8.1 VR Cross-compatibility
3.8.2 Avatar Eyeball Movement
3.8.3 Avatar Lip Tracking
3.8.4 Object Interaction
3.8.5 Notepad Improvements
References
4 A Tool for Syntactic Dependency Analysis on the Web Stack
4.1 Introduction
4.2 Background
4.3 Research Questions
4.4 Tool Design
4.5 Tool Implementation
4.6 Tool Presentation
4.7 Validation
4.8 Conclusions
References
5 Using Software for Computational Fluid Dynamics and Molecular Dynamics
5.1 Introduction
5.2 Comparison of Molecular Dynamics Methods
5.3 Computational Fluid Dynamics
5.4 Free Energy
5.5 Conclusion and Future Work
References
6 Blended Modeling Applied to the Portable Test and Stimulus Standard
6.1 Introduction
6.2 Background and Related Work
6.2.1 Blended Modeling
6.2.2 Selection of PSS Use case
6.3 Blended Modeling Framework: An Overview
6.4 Implementation and PSS use case
6.4.1 Input Format
6.4.2 Generation of Graphical and Textual Notations
6.4.3 Mapping & Synchronization
6.4.4 Editor Generation
6.5 Discussion and Future Perspectives
6.6 Conclusion and Future work
References
7 An Evaluation Framework for Modeling Languages Supporting Predictable Vehicular Software Systems
7.1 Introduction
7.2 Background and Related Work
7.2.1 Rubus Component Model (RCM)
7.2.2 Amalthea Model and APP4MC
7.2.3 Related Work
7.3 The Classification Framework
7.3.1 Timing Model
7.3.1.1 Timing Properties
7.3.1.2 Timing Requirements
7.3.2 Timing Verification
7.4 Case Study
7.4.1 Evaluation Based on Modeling of Timing Properties
7.4.2 Evaluation Based on Modeling of Timing Requirements
7.4.3 Evaluation Based on Support for Timing Verification
7.5 Conclusion and Future Work
References
8 A Model-Based Approach for Quality Assessment of Insulin Infusion Pump Systems
8.1 Introduction
8.2 Model-Based Approach
8.2.1 Hardware and Software Decomposition
8.2.2 First System Refinement
8.2.2.1 Hardware Module
8.2.2.2 Software Module
8.2.3 Second System Refinement
8.3 Quality Assessment Scenarios
8.3.1 Verification of the First Refinement
8.3.2 Validation of the Second Refinement
8.4 Empirical Evaluation
8.4.1 Scoping, Modelers, and Variables
8.4.2 Procedure and Measures
8.4.3 Analysis
8.5 Web-Based Application
8.6 Conclusions and Future Work
References
9 Narrowing the Gap Between Software Engineering Teaching and Corporate Environment
9.1 Introduction
9.2 *6pt
9.3 Tailoring Scrum and IPBL
9.4 Proof of Concept
9.5 Results and Discussion
9.6 Conclusion and Future Work
References
10 API-First Design: A Survey of the State of Academia and Industry
10.1 Introduction
10.2 Background
10.3 API-First Design
10.3.1 State of Academia
10.3.2 Industry and Grey Literature
10.4 Future Research
10.5 Summary
References
Part II Data Science & Engineering
11 A Quality Dimension Analysis of Open Government Data in Undergraduate Public Funding in Brazil
11.1 Introduction
11.2 Data Quality Dimensions
11.3 Methodology
11.4 Results and Discussion of Results
11.5 Conclusion
References
12 A Survey of Real-Time ETL Process Applied to Data Warehousing Environments
12.1 Introduction
12.2 Related Work
12.3 ETL Process
12.3.1 Extraction (E)
12.3.2 Transformation (T)
12.3.3 Loading (L)
12.4 Methodology
12.5 Related Work of Real-Time ETL Process
12.6 Conclusion and Future Work
References
13 Participatory Modeling: A New Approach to Model Graph-Oriented Databases
13.1 Introduction
13.2 Related Works
13.3 Participatory Modeling
13.4 Experiments
13.4.1 Modeling the Movies Database According to the Participatory Modeling
13.5 Conclusion
References
14 Graph-Based Hierarchical Record Clustering for Unsupervised Entity Resolution
14.1 Introduction
14.1.1 Contribution
14.2 Related Work
14.2.1 Token-Based Graph Entity Resolution
14.2.2 Record-Record Similarity-Based Graph Entity Resolution
14.2.3 Hybrid Graph-Based Entity Resolution
14.3 Method
14.3.1 Framework
14.3.1.1 Merged Data File
14.3.1.2 Parsing
14.3.1.3 Blocking
14.3.1.4 Pairwise Matching
14.3.1.5 Graph-Based Hierarchical Record Clustering (GDWM)
14.3.1.6 Canonicalization
14.4 Experiments
14.4.1 Dataset
14.4.2 Experimental Setup
14.4.3 Evaluation
14.5 Discussion
14.6 Conclusion
References
15 Semantic-MDBScan: An Approach to Assign a SemanticInterpretation to Behavior Changes Detected in Data Stream Scenarios
15.1 Introduction
15.2 Theoretical Foundations
15.2.1 Change Detection in Data Stream
15.2.2 M-DBScan
15.2.2.1 Novelty Detection
15.2.2.2 Behavior Change Detection
15.3 Related Works
15.4 Semantic-MDBScan
15.5 Experiments and Results
15.5.1 Test Datasets
15.5.2 Settings in the Experiments
15.5.3 Results
15.6 Conclusions and Future Works
References
16 A Study on Usability Assessment of Educational Systems
16.1 Introduction
16.2 Related Work
16.3 Review Methodology
16.4 Results From the Review
16.4.1 Approaches Aimed at Assessing the Usability of Educational Systems
16.4.2 Approaches Aimed at Evaluating the Usability of Mobile Learning Systems
16.4.3 Approaches Aimed at Assessing the Accessibility of E-Learning Systems
16.5 Discussion
16.6 Conclusion
A.1 Appendices
A.1.1 Appendix I
A.1.2 Appendix II
References
Part III Cybersecurity
17 Gesturing withSmart Wearables: An Alternate WaytoUser Authentication
17.1 Introduction and Background
17.2 Data Collection
17.3 Data Processing and Experimental Setups
17.3.1 Pre-processing of Data
17.3.2 Error Computation
17.4 Experimental Results
17.4.1 Additional Synchronization
17.4.2 Highlights of the Outcomes
17.5 Conclusion and Future Works
References
18 Software Optimization of Rijndael256 for Modern x86-64 Platforms
18.1 Introduction
18.2 Background and Notation
18.2.1 AES
18.2.2 AES-NI
18.2.3 Vector AES-NI
18.2.4 Rijndael256
18.3 Our Implementations
18.4 Rijndael256: The Best Possible Achievable Throughput
18.5 Experimental Setup
18.5.1 Our Optimized Code
18.5.2 Measurements Methodology
18.6 Performance Results
18.7 Conclusion
References
19 Cybersecurity Ethics Education: A Curriculum Proposal
19.1 Introduction
19.2 Background
19.3 Curriculum Proposal
19.3.1 Rationale
19.3.2 Course Description
19.3.3 Course Learning Outcomes
19.3.4 Course Topics
19.3.5 Assessment Methods
19.4 Ethical Mentoring
19.5 Conclusions
References
20 Performance Evaluation of Online Website Safeguarding Tools Against Phishing Attacks; a Comparative Assessment
20.1 Introduction
20.2 Attack Description
20.2.1 Send Phishing Emails to the Users
20.2.2 Redirect the Users to Malicious Websites
20.2.3 Request Credentials from Users
20.2.4 Forward Credentials to the Attacker
20.3 Literature Review
20.4 Phishing Detection Tools
20.5 Evaluation Procedure
20.5.1 Dataset Description
20.5.2 Evaluation Metrics
20.6 Testing Results and Discussion
20.7 Conclusion and Future Work
References
Part IV Blockchain Technology
21 Blockchain Based Trust for the Internet of Things: A Review
21.1 Introduction
21.1.1 Problem Description and Motivation
21.2 Background and Related Works
21.2.1 Background
21.2.2 Related Works
21.3 Proposed Taxonomy Scheme
21.4 Assessment of Blockchain Based Trust Approaches in IoT
21.5 Conclusion
References
22 The Use ofBlockchain Technology inElectronic HealthRecord Management: An Analysis ofState oftheArt andPractice
22.1 Introduction
22.2 Background
22.3 Blockchain-Based EHR Management Analysis
22.3.1 An Analysis of the State of the Art
22.3.2 An Analysis of the State of the Practice
22.4 Discussion
22.5 Final Considerations
References
23 Blockchain for Security and Privacy of Healthcare Systems: A Protocol for Systematic Literature Review
23.1 Introduction
23.1.1 Rationale
23.1.2 Objectives
23.2 Systematic Literature Method
23.2.1 Eligibility Criteria
23.2.2 Information Sources
23.2.3 Search Strategy
23.2.4 Study Record
23.2.4.1 Data Management
23.2.4.2 Selection Process
23.2.5 Data Collection Process
23.2.6 Data Items
23.2.7 Outcomes and Prioritization
23.2.7.1 Primary Outcomes
23.2.7.2 Secondary Outcomes
23.2.8 Risk of Bias Individual Studies
23.2.9 Data Synthesis
23.2.10 Confidence in Cumulative Estimate and Meta-Bias
23.3 Conclusion
References
24 Single Sign-On (SSO) Fingerprint Authentication Using Blockchain
24.1 Introduction
24.2 Backend of OUR Proposed Web App
24.3 Frontend of the Proposed Web App
24.4 Conclusion and Future Work
References
Part V Health Informatics
25 A Detection Method for Early-Stage Colorectal Cancer Using Dual-Tree Complex Wavelet Packet Transform
25.1 Introduction
25.2 2D Complex Discrete Wavelet Packet Transform
25.2.1 2D Complex Discrete Wavelet Transform
25.2.2 Dual-Tree Complex Wavelet Packet Transform
25.3 Proposed Method
25.3.1 Preprocessing for Endoscopic Image
25.3.2 Preliminary Experiment
25.3.3 Proposed Detection Method
25.4 Experimental Results
25.4.1 Classification Experiment
25.4.2 Comparison Experiment
25.5 Conclusion
References
26 Visualizing 3D Human Organs for Medical Training
26.1 Introduction
26.2 Research Methodology
26.2.1 Overview
26.2.2 Experiment Setup
26.2.3 Experiment Implementation
26.3 Experimental Results
26.3.1 Results for Our 3D Models
26.3.2 Analysis and Discussion
26.4 Conclusions and Future Work
A.1 Appendices
A.1.1 Appendix A: The 3D Models Used in the Survey
A.1.2 Appendix B: The 2D Images Used in the Survey
References
27 An Information Management System for the COVID-19 Pandemic Using Blockchain
27.1 Introduction
27.2 Tailoring the Scrum
27.3 Background
27.3.1 The STEPES-BD Architecture
27.3.2 The Infrastructure for the STEPES-BD Project
27.3.3 The Software Application
27.3.4 The Data Interface
27.3.5 Testing the Application
27.4 The Proof of Concept (PoC)
27.5 Conclusion
References
28 Machine Learning for Classification of Cancer Dataset for Gene Mutation Based Treatment
28.1 Introduction
28.2 Background
28.3 Proposed System Architecture
28.4 Experimental Results
28.5 Conclusion and Future Work
References
Part VI Machine Learning
29 Performance Comparison Between Deep Learning and Machine Learning Models for Gene Mutation-Based Text Classification of Cancer
29.1 Introduction
29.2 Background
29.3 Proposed System Architectures
29.3.1 Deep Learning Neural Network Models
29.3.2 Machine Learning Algorithms
29.4 Experimental Results
29.5 Conclusion and Future Work
References
30 Stock Backtesting Engine Using Pairs Trading
30.1 Introduction
30.2 Background Work
30.3 Our Proposed System
30.4 Results Based on Yahho Finance Data
30.5 Conclusion and Future Work
References
31 Classifying Sincerity Using Machine Learning
31.1 Introduction
31.2 Background Work
31.3 Machine Learning Models Applied
31.4 Results of Various Machine Learning Algorithms
31.5 Conclusion and Future Work
References
32 Recommendation System Using MixPMF
32.1 Introduction
32.2 Background Work
32.3 Our Proposed MixPMF Model
32.4 Model Training and Evaluation
32.5 Conclusion and Future Work
References
33 Abstractive Text Summarization Using Machine Learning
33.1 Introduction
33.2 Background and Related Work
33.3 Implemented Text Summarization Algorithms
33.3.1 Extractive Text Summarization Using NLTK
33.3.2 Extractive Text Summarization Using TextRank
33.3.3 Abstractive Text Summarization Using Seq-to-Seq
33.4 Simulations and Results
33.5 Conclusion and Future Work
References
34 Intelligent System for Detection and Identification of Ground Anomalies for Rescue
34.1 Introduction
34.2 Related Works
34.3 Methodology
34.4 Provisional Results
34.5 Conclusion
Reference
Part VII Human-Computer Interaction
35 An Application for Interaction Comparison Between Virtual Hands and Virtual Reality Controllers
35.1 Introduction
35.2 Background and Related Works
35.2.1 HCI VE Considerations
35.2.2 VE Interaction Design
35.2.3 Virtual Input Devices
35.2.4 VR Gloves
35.3 Methods
35.3.1 Virtual Input Usability Factors
35.3.2 Detailed Use Cases
35.4 Implementation
35.4.1 Stage 1: Menu Interaction
35.4.2 Stage 2: Virtualized Everyday Objects
35.4.3 Stage 3: Rigidbody Physics Object Interactions
35.5 Conclusion
35.6 Future Work
References
36 LDAT: A LIDAR Data Analysis and Visualization Tool
36.1 Introduction
36.2 Background and Related Work
36.3 Software Design
36.3.1 System Architecture
36.3.2 Data Workflow
36.3.3 Application Design
36.4 Prototype
36.4.1 User Interface
36.4.2 Data Visualizations
36.5 User Study and Results
36.5.1 User Study
36.5.2 Results
36.6 Conclusion and Future Work
References
References
37 Social Media User Study
37.1 Introduction
37.2 Methodology
37.2.1 Participants
37.2.2 Apparatus
37.2.3 Procedure
37.2.4 Tasks
37.2.5 Design
37.3 Results and Discussion
37.3.1 Data Analysis
37.3.2 Questionnaire Analysis
37.4 Conclusions and Future Work
References
38 Software Interfaces for New Vehicle Operating Cost Models Used in Economic Analysis of Transportation Investments: A User Study
38.1 Introduction
38.2 Vehicle Cost Operation Models
38.2.1 Background on VOC Research
38.2.2 Vehicle Cost Operation Models
38.2.3 VOC Models Application and Challenges
38.3 Excel and Web Interfaces
38.3.1 VOC Models
38.3.2 Excel Interface
38.3.3 Website Interface
38.4 User Study
38.4.1 Participants
38.4.2 Technology
38.4.3 Procedure
38.4.4 Tasks
38.4.5 Design
38.5 Results
38.5.1 Results Obtained from the Entry Questionnaire
38.5.2 Results Obtained from the Tasks Performed by the Participants
38.5.3 Results Obtained from the Exit Questionnaire
38.6 Discussion
38.7 Conclusions and Future Work
References
39 Microservice-Based System for Environmental Science SoftwareApplications
39.1 Introduction
39.2 Background and Related Works
39.2.1 Nevada Research Data Center
39.2.2 Microservice Architecture
39.2.3 DIMMER Smart City Platform
39.2.4 Generic Service Infrastructure for PEIS
39.2.5 OceanTEA: Exploring Ocean-Derived Climate Data
39.3 Software Specification
39.3.1 High Level Design
39.3.2 Technology Utilized
39.4 Use Cases
39.4.1 SEPHAS Lysimeter Visualization
39.4.2 NRDC Quality Assurance Application
39.4.3 Conflict Management
39.4.4 Near Real-Time Autonomous Quality Control
39.5 Discussion
39.6 Conclusions and Future Work
39.6.1 Conclusion
39.6.2 Future Work
References
Part VIII Networks
40 Semantic Interoperability in the Internet of Things: A Systematic Literature Review
40.1 Introduction
40.2 Method
40.2.1 Planning the Review
40.2.2 Conducting the Review
40.2.3 Reporting the Review
40.3 Results
40.3.1 Research Question 1
40.3.2 Research Question 2
40.3.3 Research Question 3
40.4 Conclusion
References
41 IoT Machine Learning Based Parking Management System with Anticipated Prediction of Available Parking Spots
41.1 Introduction
41.2 Related Work
41.3 Proposed System
41.3.1 Mobile App
41.3.2 Database
41.3.3 Machine Learning Model
41.3.3.1 Purpose
41.3.3.2 Dataset
41.3.3.3 Model Parameters
41.3.3.4 Model Architecture and Training
41.3.3.5 Model Results
41.3.4 Backend Server
41.3.5 Geofences
41.3.6 Model Prediction Consequences
41.3.7 Time Prediction
41.3.7.1 Multi-floor Garages
41.3.7.2 Dynamically Calculating Speed
41.3.7.3 Accounting for Uncertainties
41.4 Conclusions
References
42 Channel State Information Spectrum Gap Filling Using Shallow Neural Networks
42.1 Introduction
42.2 Related Work
42.3 Overview of CSIFill
42.4 Details of CSIFill
42.4.1 Training Algorithms
42.4.2 Network Configuration
42.4.2.1 Configuration for Bounded Estimation
42.4.2.2 Configuration for Unbounded Estimation
42.4.3 Training the Network
42.4.4 Testing the Network
42.5 Evaluation
42.5.1 Data Collection and Pre-processing
42.5.2 Evaluation Process
42.5.3 MSE Performance
42.5.3.1 Bounded Estimation Accuracy
42.5.3.2 Unbounded Estimation Accuracy
42.6 Conclusion and Future Work
References
Part IX Potpourri
43 Unveiling a Novel Corporate Structure in World-Class Business, Merging Digital-Physical Environment in Hyper Famili Incorporation
43.1 Introduction
43.2 Literature
43.2.1 World-Class Business Ecosystem Model
43.2.2 Intelligent Seamless Ecosystem
43.2.3 Hyper Famili Incorporation
43.3 The Physical-Digital Boundary
43.3.1 Training; the Immersive Learning as a Mandatory Technology in New Ecosystems
43.3.2 Benefits of Immersive Learning
43.4 Our Nearest Mega-Retail Rivals Strategic Ecosystem Structure Plan
43.5 Brand New Warehouse and Inventory Management Model for Hyper Famili Ecosystem
43.6 Organizational Structure
43.7 Comparative Study in Hyper Famili
43.8 Biomatrix Corporate Structure
43.9 Conclusion
References
44 Developing an Affective Audio Toolbox for Audio Post-production
44.1 Introduction
44.2 Background
44.3 Methodology
44.4 Results
44.5 Analysis and Discussion
44.5.1 Sound as an Affective Device
44.5.2 Sound as a Narrative Device
44.5.3 Sound Design Technique
44.5.4 Mixing Technique
44.5.5 Role of a Sound Editor/Mixer
44.5.6 Director's Input to Sound
44.6 Themes in Action
44.6.1 Blade Runner (1982) & Blade Runner 2049 (2017)
44.6.2 Once Upon a Time… in Hollywood (2019)
44.7 Working with Themes
44.8 Conclusion and Future Work
References
45 Boundary Approximation and External Visibility
45.1 Introduction
45.2 Review of Polygonal Approximation
45.3 Visibility and Boundary Approximation
45.4 Discussions
References
46 Detection of Strictly L3-Live Structures by Structural Analysis of General Petri Net Using SAT-Solver
46.1 Introduction
46.2 Petri Nets
46.2.1 Place/Transition Net
46.2.2 Incidence Matrix
46.2.3 Firing Condition
46.2.4 Structural Properties
46.2.5 Liveness
46.3 Proposal of Srictly L3-Live Structure
46.4 Structural Analysis
46.4.1 Solving structural properties using SAT-solver
46.4.2 Detection of CircuitBreaker
46.4.3 Detection of L3-circuit
46.4.4 Detection of k-place
46.5 Example of Structural Analysis
46.6 Conclusions and Future Work
References
47 Space Abstraction and of PetriNets Using the Submarking Method Quasi-home States
47.1 Introduction
47.2 Petri Nets
47.2.1 Overview of Petri Nets
47.2.2 Reachability Analysis and Reachability Coverability Analysis
47.3 Submarking and Qusai-home State Proposals
47.3.1 Qusai-home State
47.3.2 Submarking
47.4 Application Examples
47.4.1 Application Examples 1 (Removing the Insert-Reject Loop)
47.4.2 Application Examples 1 (Abstracting by Inventory Count)
47.4.3 Application Examples 3 (Non-bounded Place Abstraction)
47.5 Performance Comparison
References
Index
Recommend Papers
- Author / Uploaded
- Shahram Latifi (editor)
File loading please wait...
Citation preview
Advances in Intelligent Systems and Computing 1421
Shahram Latifi Editor
ITNG 2022 19th International Conference on Information Technology-New Generations
Advances in Intelligent Systems and Computing Volume 1421 Series Editor Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland
The series “Advances in Intelligent Systems and Computing” contains publications on theory, applications, and design methods of Intelligent Systems and Intelligent Computing. Virtually all disciplines such as engineering, natural sciences, computer and information science, ICT, economics, business, e-commerce, environment, healthcare, life science are covered. The list of topics spans all the areas of modern intelligent systems and computing such as: computational intelligence, soft computing including neural networks, fuzzy systems, evolutionary computing and the fusion of these paradigms, social intelligence, ambient intelligence, computational neuroscience, artificial life, virtual worlds and society, cognitive science and systems, Perception and Vision, DNA and immune based systems, self-organizing and adaptive systems, eLearning and teaching, human-centered and human-centric computing, recommender systems, intelligent control, robotics and mechatronics including human-machine teaming, knowledgebased paradigms, learning paradigms, machine ethics, intelligent data analysis, knowledge management, intelligent agents, intelligent decision making and support, intelligent network security, trust management, interactive entertainment, Web intelligence and multimedia. The publications within “Advances in Intelligent Systems and Computing” are primarily proceedings of important conferences, symposia and congresses. They cover significant recent developments in the field, both of a foundational and applicable character. An important characteristic feature of the series is the short publication time and world-wide distribution. This permits a rapid and broad dissemination of research results. Indexed by DBLP, INSPEC, WTI Frankfurt eG, zbMATH, Japanese Science and Technology Agency (JST). All books published in the series are submitted for consideration in Web of Science. For proposals from Asia please contact Aninda Bose ([email protected]). Advisory Board Nikhil R. Pal, Indian Statistical Institute, Kolkata, India Rafael Bello Perez, Faculty of Mathematics, Physics and Computing, Universidad Central de Las Villas, Santa Clara, Cuba Emilio S. Corchado, University of Salamanca, Salamanca, Spain Hani Hagras, School of Computer Science and Electronic Engineering, University of Essex, Colchester, UK László T. Kóczy, Department of Automation, Széchenyi István University, Gyor, Hungary Vladik Kreinovich, Department of Computer Science, University of Texas at El Paso, El Paso, TX, USA Chin-Teng Lin, Department of Electrical Engineering, National Chiao Tung University, Hsinchu, Taiwan Jie Lu, Faculty of Engineering and Information Technology, University of Technology Sydney, Sydney, NSW, Australia Patricia Melin, Graduate Program of Computer Science, Tijuana Institute of Technology, Tijuana, Mexico Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro, Rio de Janeiro, Brazil Ngoc Thanh Nguyen, Faculty of Computer Science and Management, Wrocław University of Technology, Wrocław, Poland Jun Wang, Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, Hong Kong
More information about this series at https://link.springer.com/bookseries/11156
Shahram Latifi Editor
ITNG 2022 19th International Conference on Information Technology-New Generations
Editor Shahram Latifi Department of Electrical and Computer Engineering University of Nevada Las Vegas, NV, USA
ISSN 2194-5357 ISSN 2194-5365 (electronic) Advances in Intelligent Systems and Computing ISBN 978-3-030-97651-4 ISBN 978-3-030-97652-1 (eBook) https://doi.org/10.1007/978-3-030-97652-1 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Contents
Chair’s Message . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ix
Part I Software Engineering 1
Cross-Platform Blended Modelling with JetBrains MPS and Eclipse Modeling Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Malvina Latifaj, Hilal Taha, Federico Ciccozzi, and Antonio Cicchetti
3
2
A Conceptual Framework for Software Modeling of Automation Systems . . . Mohammad Ashjaei, Alessio Bucaioni, and Saad Mubeen
3
Virtual Reality Multiplayer Interaction and Medical Patient Handoff Training and Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Christopher Lewis, Daniel Enriquez, Lucas Calabrese, Yifan Zhang, Steven J. Anbro, Ramona A. Houmanfar, Laura H. Crosswell, Michelle J. Rebaleati, Luka A. Starmer, and Frederick C. Harris, Jr.
4
A Tool for Syntactic Dependency Analysis on the Web Stack . . . . . . . . . . . . . . . Manit Singh Kalsi, Kevin A. Gary, Vasu Gupta, and Suddhasvatta Das
25
5
Using Software for Computational Fluid Dynamics and Molecular Dynamics Jeena Shetti, Stefan Pickl, Doina Bein, and Marian Sorin Nistor
35
6
Blended Modeling Applied to the Portable Test and Stimulus Standard . . . . . Muhammad Waseem Anwar, Malvina Latifaj, and Federico Ciccozzi
39
7
An Evaluation Framework for Modeling Languages Supporting Predictable Vehicular Software Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Enxhi Ferko, Igli Jasharllari, Alessio Bucaioni, Mohammad Ashjaei, and Saad Mubeen
8
9
A Model-Based Approach for Quality Assessment of Insulin Infusion Pump Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tássio Fernandes Costa, Álvaro Sobrinho, Lenardo Chaves e Silva, Leandro Dias da Silva, and Angelo Perkusich Narrowing the Gap Between Software Engineering Teaching and Corporate Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Marcelo A. M. da Conceicao, Oswaldo S. C. Neto, Andre B. Baccarin, Luan H. S. Dantas, Joao P. S. Mendes, Vinicius P. Lippi, Gildarcio S. Gonçalves, Adilson M. Da Cunha, Luiz A. Vieira Dias, Johnny C. Marques, and Paulo M. Tasinaffo
10 API-First Design: A Survey of the State of Academia and Industry . . . . . . . . . Nicole Beaulieu, Sergiu M. Dascalu, and Emily Hand
11
47
57
65
73
v
vi
Contents
Part II Data Science & Engineering 11
12
13
14
15
16
A Quality Dimension Analysis of Open Government Data in Undergraduate Public Funding in Brazil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Marcelo Moreira West and Glauco de Figueiredo Carneiro
83
A Survey of Real-Time ETL Process Applied to Data Warehousing Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Flávio de Assis Vilela and Ricardo Rodrigues Ciferri
91
Participatory Modeling: A New Approach to Model Graph-Oriented Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Luis A. Neumann, Enzo Seraphim, Otávio A. O. Carpinteiro, and Edmilson M. Moreira
97
Graph-Based Hierarchical Record Clustering for Unsupervised Entity Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Islam Akef Ebeid, John R. Talburt, and Md Abdus Salam Siddique
107
Semantic-MDBScan: An Approach to Assign a Semantic Interpretation to Behavior Changes Detected in Data Stream Scenarios . . . . . Eldane Vieira Júnior, Rita Maria Silva Julia, and Elaine Ribeiro Faria
119
A Study on Usability Assessment of Educational Systems . . . . . . . . . . . . . . . . . . Heber Miranda Floriano, Mario Jino, and Ferrucio de Franco Rosa
129
Part III Cybersecurity 17
Gesturing with Smart Wearables: An Alternate Way to User Authentication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Khandaker Abir Rahman, Avishek Mukherjee, and Kristina Mullen
137
18
Software Optimization of Rijndael256 for Modern x86-64 Platforms . . . . . . . . . Nir Drucker and Shay Gueron
147
19
Cybersecurity Ethics Education: A Curriculum Proposal . . . . . . . . . . . . . . . . . . Ping Wang
155
20
Performance Evaluation of Online Website Safeguarding Tools Against Phishing Attacks; a Comparative Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rama Al-Share, Fatima Abu-Akleek, Ahmed S. Shatnawi, and Eyad Taqieddin
161
Part IV Blockchain Technology 21
Blockchain Based Trust for the Internet of Things: A Review . . . . . . . . . . . . . . Dina Shehada, Maryam Amour, Suadad Muammar, and Amjad Gawanmeh
22
The Use of Blockchain Technology in Electronic Health Record Management: An Analysis of State of the Art and Practice . . . . . . . . . . Henrique Couto, André Araújo, Rendrikson Soares, and Gabriel Rodrigues
179
Blockchain for Security and Privacy of Healthcare Systems: A Protocol for Systematic Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Saadia Azemour, Meryeme Ayache, Hanane El Bakkali, and Amjad Gawanmeh
187
23
24
Single Sign-On (SSO) Fingerprint Authentication Using Blockchain . . . . . . . . Abhijeet Thakurdesai, Marian Sorin Nistor, Doina Bein, Stefan Pickl, and Wolfgang Bein
171
195
Contents
vii
Part V Health Informatics 25 A Detection Method for Early-Stage Colorectal Cancer Using Dual-Tree Complex Wavelet Packet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Daigo Takano and Teruya Minamoto 26 Visualizing 3D Human Organs for Medical Training . . . . . . . . . . . . . . . . . . . . . . Joshua Chen, Paul J. Cuaresma, Jennifer S. Chen, Fangyang Shen, and Yun Tian 27 An Information Management System for the COVID-19 Pandemic Using Blockchain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Marcelo Alexandre M. da Conceicao, Oswaldo S. C. Neto, Andre B. Baccarin, Luan H. S. Dantas, Joao P. S. Mendes, Vinicius P. Lippi, Gildarcio S. Gonçalves, Adilson M. Da Cunha, Luiz A. Vieira Dias, Johnny C. Marques, and Paulo M. Tasinaffo 28 Machine Learning for Classification of Cancer Dataset for Gene Mutation Based Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jai Santosh Mandava, Abhishek Verma, Fulya Kocaman, Marian Sorin Nistor, Doina Bein, and Stefan Pickl
205 211
219
229
Part VI Machine Learning 29 Performance Comparison Between Deep Learning and Machine Learning Models for Gene Mutation-Based Text Classification of Cancer . . . . . . . . . . . . Fulya Kocaman, Stefan Pickl, Doina Bein, and Marian Sorin Nistor
237
30 Stock Backtesting Engine Using Pairs Trading . . . . . . . . . . . . . . . . . . . . . . . . . . . Rahul Chauhan, Marian Sorin Nistor, Doina Bein, Stefan Pickl, and Wolfgang Bein
245
31 Classifying Sincerity Using Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . Rachana Chittari, Marian Sorin Nistor, Doina Bein, Stefan Pickl, and Abhishek Verma
255
32 Recommendation System Using MixPMF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rohit Gund, James Andro-Vasko, Doina Bein, and Wolfgang Bein
263
33 Abstractive Text Summarization Using Machine Learning . . . . . . . . . . . . . . . . . Aditya Dingare, Doina Bein, Wolfgang Bein, and Abhishek Verma
269
34 Intelligent System for Detection and Identification of Ground Anomalies for Rescue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Antonio Dantas, Leandro Diniz, Maurício Almeida, Ella Olsson, Peter Funk, Rickard Sohlberg, and Alexandre Ramos
277
Part VII Human-Computer Interaction 35 An Application for Interaction Comparison Between Virtual Hands and Virtual Reality Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Daniel Enriquez, Christopher Lewis, Sergiu M. Dascalu, and Frederick C. Harris, Jr.
285
36 LDAT: A LIDAR Data Analysis and Visualization Tool . . . . . . . . . . . . . . . . . . . . Andrew Muñoz, Chase Carthen, Vinh Le, Scotty D. Strachan, Sergiu M. Dascalu, and Frederick C. Harris, Jr.
293
37 Social Media User Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Autumn Cuellar, Yifan Zhang, Sergiu M. Dascalu, and Frederick C. Harris, Jr.
303
viii
38
39
Contents
Software Interfaces for New Vehicle Operating Cost Models Used in Economic Analysis of Transportation Investments: A User Study . . . . . . . . . . . . . . . . . . . . Arjun V. Gopinath, Hudson Lynam, Rami Chkaiban, Elie Hajj, and Sergiu M. Dascalu Microservice-Based System for Environmental Science Software Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vinh Le, Connor Scully Allison, Mitchell Martinez, Sergiu M. Dascalu, Frederick C. Harris, Jr., Scotty D. Strachan, and Eric Fritzinger
311
321
Part VIII Networks 40
41
42
Semantic Interoperability in the Internet of Things: A Systematic Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Pedro Lopes de Souza, Wanderley Lopes de Souza, and Ricardo Rodrigues Ciferri
333
IoT Machine Learning Based Parking Management System with Anticipated Prediction of Available Parking Spots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Grzegorz Chmaj and Michael Lazeroff
341
Channel State Information Spectrum Gap Filling Using Shallow Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Avishek Mukherjee, Beata Hejno, and Manish Osti
351
Part IX Potpourri 43
Unveiling a Novel Corporate Structure in World-Class Business, Merging Digital-Physical Environment in Hyper Famili Incorporation . . . . . . . . . . . . . . Mohammad Khakzadeh, Fatemeh Saghafi, Seyed Milad Seyed Javadein, Mohammad Hossein Asmaie, and Masoud Matbou Saleh
363
44
Developing an Affective Audio Toolbox for Audio Post-production . . . . . . . . . . Harrison Ridley, Stuart Cunningham, and Richard Picking
371
45
Boundary Approximation and External Visibility . . . . . . . . . . . . . . . . . . . . . . . . . Laxmi Gewali and Samridhi Jha
379
46
Detection of Strictly L3-Live Structures by Structural Analysis of General Petri Net Using SAT-Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuta Yoshizawa and Katsumi Wasaki
387
Space Abstraction and of PetriNets Using the Submarking Method Quasi-home States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tomoki Miura and Katsumi Wasaki
393
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
399
47
Chair’s Message
Welcome to the 19th International Conference on Information Technology – New Generations – ITNG 2022. Due to the continued global pandemic and the problems associated with traveling and in-person meeting, we are running the conference virtually. We hope and trust that in the following year, we will be able to meet in person when the audience feel safer and more comfortable to travel and participate in this event. The ITNG 2022 attracted quality submissions globally. The papers were reviewed for their technical soundness, originality, clarity, and relevance to the conference. The conference enjoyed expert opinion of over 40 author and non-author scientists who participated in the review process. Each paper was reviewed by at least two independent reviewers. At the end, 48 papers were accepted for presentation and publication in the ITNG 2022 program. The chapters in this book address the most recent advances in such areas as machine learning, big data analytics, cybersecurity, blockchain technology, data mining, e-health, IoT & CPS, software engineering and social computing. In addition to technical presentations by the authors, the conference features two keynote speakers on Monday and Tuesday. Several people contributed to the success of this year’s conference by organizing technical tracks. Dr. Doina Bein served in the capacity of conference vice chair. We benefited from the professional and timely services of major track organizers and associate editors, namely Drs. Azita Bahrami, Christian Barria, Doina Bein, Allesio Bucaioni, Glauco Carneiro, Sergiu Dascalu, Luiz Alberto Vieira Dias, Fred Harris, Ray Hashemi, Kashif Saleem, Fangyan Shen, David Cordero Vidal, and Hossein Zare. Others who were responsible for solicitation, review, and handling the papers submitted to their respective tracks/sessions include Drs. Wolfgang Bein, Poonam Dharam, and Mei Yang. The help and support of Springer in preparing the ITNG proceedings is specially appreciated. Many thanks are due to Michael Luby, Senior Editor and Supervisor of Publications, and Brian Halm, Production Editor at Springer, for the timely handling of our publication order. We also appreciate the efforts made by the Springer Project Coordinator, Olivia Ramya Chitranja. Olivia spent much time looking very closely at revised chapters to make sure they are formatted correctly according to the publisher’s guidelines. ix
x
Chair’s Message
We also thank the technical assistance of Prabhas Kumra in setting up the conference program and Zoom communication for us. Finally, the great efforts of the conference secretary, Ms. Mary Roberts, who dealt with the day-to-day conference affairs, including timely handling of emails, are acknowledged. I hope that you all enjoy the ITNG 2022 program and find it technically and socially fulfilling. The ITNG General Chair
Shahram Latifi
ITNG 2022 Reviewers
Andro-Vasko, James Abbas, Haider Alahmad, Yanal Alasady, Majida Alwadi, Ali Bahrami, Azita Barría, Cristian Bein, Doina Bein, Wolfgang Cagnin, Maria Carneiro, Glauco Cirstea, Silvia Collazos, Cesar Cordero, David Cunha, Adilson Marques da Cunningham, Stuart Dascalu, Sergiu David, José Dharam, Poonam Dias, Luis Alberta Vieira Garcia, Vinicius Gawanmeh, Amjad Gawanmeh, Amjad Grout, Vic Harris, Frederick Hashemi, Ray Hazzazi, Noha
Latifi, Shahram Le, Vinh Liu, Siming Machado, Ivan Marques, Johnny Mascarenhas, Ana Mialaret, Lineu Fernando Stegr Mohammadi-Koushki, Negar Mohd-Ali, Nursabillilah Monteiro, Miguel Muhanna, Muhanna Nguyen, Tin Quiroz, Juan Redei, Alex Resende, Antonio Saleem, Kashif Scully-Allison, Connor Sharma, Sharad Shen, Fangyang Silva, Paulo Caetano Soares, Michel Suzana, Rita Trausan-Matu, Stefan Vieira-Dias, Luiz-Alberto Yang, Mei Zare, Hossein
xi
Part I Software Engineering
1
Cross-Platform Blended Modelling with JetBrains MPS and Eclipse Modeling Framework Malvina Latifaj, Hilal Taha, Federico Ciccozzi, and Antonio Cicchetti
Abstract
Modelling tools traditionally focus on one specific editing notation (such as text, diagrams, tables or forms), providing additional visualisation-only notations. For softwareintensive systems with heterogeneous components and entailing different domain-specific aspects and different stakeholders, one editing notation is too little and voids many modelling benefits. Seamless blended modelling, which allows stakeholders to freely choose and switch between graphical and textual notations, can greatly contribute to increase productivity as well as decrease costs and time to market. In this paper we describe our work in bridging two powerful (meta) modelling platforms: Eclipse Modeling Framework, for the definition of treebased and graphical DSMLs and models conforming to them, and JetBrains MPS, for the description of textual DSMLs and the projectional manipulation of textual models via multiple views. The possibility to visualise and edit the same information in these two platforms, otherwise disjoint, can greatly boost communication between stakeholders, who can freely select their preferred notation or switch from one to the other at any time. Keywords
Blended modelling · Graphical modelling · Textual modelling · Model transformations · Language engineering · Model-driven engineering · Language workbenches · MPS · EMF · Ecore M. Latifaj () · H. Taha · F. Ciccozzi · A. Cicchetti School of Innovation, Design and Engineering, Malardalen University, Vasteras, Sweden e-mail: [email protected]; [email protected]; [email protected]; [email protected]
1.1
Introduction
The complexity of modern software systems is continuously growing. It is enough to mention the increasing trend of smartifying applications by augmenting them with self-adaptation and customisation features. With this trend, software tends to cross-cut and integrate multidisciplinary domains [4], posing the problem of keeping consistent the properties of the system from multiple viewpoints [8]. In any non-trivial development scenario, consistency management is impracticable if performed at the implementation code level of abstraction. Or more precisely, it could be possible, but the consequences of detected malfunctions are typically expensive analyses to understand where a problem originates and how it could be fixed [13]. Model-Driven Engineering (MDE) methodology proposes to use models as the main development artefacts [12]. In this way, it is possible to reduce the complexity of systems development, and notably validate the integration earlier in the process thanks to a higher-level of abstraction view of the result; in turn, these analyses potentially save the costs of modifications because they can significantly anticipate the implementation phases. Moreover, when dealing with multidisciplinary domains, MDE state-ofthe-art recommends to use Domain-Specific Modelling Languages (DSMLs): as the name suggests, DSMLs are languages tailored for a specific domain, hence providing experts with modelling concepts closer to their knowledge areas [13]. The set of modelling concepts included in a DSML is defined by means of a metamodel, that can be considered as a grammar defining the set of legal productions for a language. In this respect, language engineers usually leverage workbenches to create DSMLs [10]; in fact, based on metamodel specifications language workbenches
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_1
3
4
M. Latifaj et al.
provide effective features for a DSML provision, notably the automatic derivation of model management processes, conformance checks, the support for the definition of custom editors, and so forth. Depending on the technical space [3], the preferred underlying implementation technologies, etc., there exists a multitude of alternative language workbenches. Even if from a conceptual perspective they are all based on similar metamodel specifications, only in rare cases it is possible to exchange metamodels, and corresponding models, between workbenches. This paper discusses our efforts in bridging two language workbenches, JetBrains MPS (later also referred to as simply MPS) [10] and the Eclipse Modeling Framework (EMF) [7]; more specifically, the work is devoted to the mapping of metamodels defined in MPS towards metamodels conforming to the EMF specification language, namely Ecore. Technically, we firstly contribute with an Ecore language specification for MPS; based on it, users can create metamodels by using the MPS language workbench features and possibly create models conforming to such metamodels. Alternatively, we contribute with a transformation for mapping the metamodels defined in MPS as Ecore metamodels usable in EMF. In this latter scenario, users can leverage EMF plug-ins e.g. to generate default tree editors, create a custom concrete syntax, use the metamodel as part of a model transformation chain, and so forth. The paper is organised as follows: Sect. 1.2 describes the motivation and intended contribution, while Sect. 1.3 compares related research in the area with respect to this contribution. The details of the developed solution and its evaluation are provided in Sects. 1.4 and 1.5 respectively. Eventually, the paper discusses concluding remarks and draws future investigation directions in Sect. 1.6.
1.2
aspects and different types of stakeholders, mutual exclusion is too restrictive and voids many of the MDE benefits [5]. Therefore, modelling tools need to enable different stakeholders to work on overlapping parts of the models using different modelling notations (e.g., graphical and textual). We have previously defined the notion of blended modelling [6] as: the activity of interacting seamlessly with a single model (i.e., abstract syntax) through multiple notations (i.e., concrete syntaxes), allowing a certain degree of temporary inconsistencies.
Seamless blended modelling, which allows stakeholders to freely choose and switch between graphical and textual notations, can greatly contribute to increase productivity as well as decrease costs and time to market, as we have preliminary shown in [2]. In this paper we describe our work in bridging two powerful (meta) modelling platforms: EMF, for the definition of tree-based and graphical DSMLs and models conforming to them, and MPS, for the description of textual DSMLs and the projectional manipulation of textual models via multiple views. The possibility to visualise and edit the same information in these two platforms, otherwise disjoint, can greatly boost communication between stakeholders, who can freely select their preferred notation or switch from one to the other at any time. The research contribution is twofold: (i) definition and implementation of a textual language in MPS for representing Ecore’s meta-metamodel; this enables the creation of Ecore-based meta-models directly in MPS, and (ii) definition and implementation of a mapping mechanism for Ecore-based meta-models created in MPS to be imported in EMF, enabling the automated exchange of these metamodels across the two platforms.
Motivation and Contribution 1.3
Modelling tools traditionally focus on one specific editing notation (such as text, diagrams, tables or forms). This limits human communication, especially across stakeholders with varying roles and expertise. Moreover, engineers may have different notation preferences; not supporting multiple notations negatively affects their throughput. Besides the limits to communication, choosing one particular kind of notation has the drawback of limiting the pool of available tools to develop and manipulate models that may be needed. For example, choosing a graphical representation limits the usability of text manipulation tools such as text-based diff/merge, which is essential for team collaboration. When tools provide support for both graphical and textual modelling, it is mostly done in a mutual exclusive manner. For systems with heterogeneous components and entailing different domain-specific
Related Work
In MPS it exists a mechanism to import and export Ecorebased languages (metamodels). Metamodels are imported as language structures in MPS [1], allowing the user to create models conforming to it and add structural features to imported metamodels. However, this does not allow to create Ecore-based metamodels in MPS, which is instead our main goal rather than simply importing them. The main difference is that we reproduce Ecore entirely in MPS, instead of transforming Ecore-based metamodels into MPS languages, which would lead to loss of conformance to the Ecore metamodel itself in case of modifications. EMFText is an approach that allows the definition of textual syntaxes for languages described in terms of Ecore. More specifically, it allows the developers to define DSMLs
1 Cross-Platform Blended Modelling with JetBrains MPS and Eclipse Modeling Framework
using a native textual syntax in EMF [9]. Being part of EMF, this approach does not provide any of the projectional benefits offered by MPS, the reason why we chose MPS as textual modelling target platform. Scheidgen [11] provides techniques to combine textual and graphical modelling by embedding textual editors into graphical ones. Users open an embedded text editor by clicking on an element within the graphical host editor. This work can enhance the effectiveness of graphical modelling with languages that partially rely on textual representations. The author focuses on adding functionalities to graphical editors by embedding textual models. Nevertheless, most benefits of the projectional approach provided by MPS would not be leverageable. Related to the switching between graphical and textual syntaxes, Wimmer et al. [14] propose to bridge Grammarware and Modelware to ease transformations of models containing both graphical and textual elements. In Grammarware, a mixed model is exported as text. In Modelware [14], a model containing graphical and textual content is transformed into a fully graphical model. Although related to our work in terms of blending graphical and textual notations, our aim is to bridge two well-established frameworks providing
Fig. 1.1 Process of bridging MPS and EMF, and evaluation
5
rather different approaches to modelling in order to exploit the benefits of both and we do it by a common language, Ecore.
1.4
Bridging MPS and EMF
To connect the textual-based world of MPS and the diagrambased world of EMF, we propose a solution that enables the textual definition of metamodels conforming to the Ecore meta-metamodel in MPS, and the exchange of these metamodels from MPS to EMF. Ecore is the language used in EMF for describing models, which are serialized and persisted as XMI files [7]. Throughout the rest of the text, we refer to metamodels created in EMF as EcoreEMF metamodels (.ecore extension), and the ones created in MPS as EcoreMPS metamodels (.sandbox.mps extension). Implementation and evaluation of this solution are carried out by following the process illustrated in Fig. 1.1. The process consists of two main steps. Step 1. Recreate the Ecore meta-metamodel, as defined in EMF, as a language (EcoreLanguage) in MPS and evaluate whether the EcoreMPS metamodel conforming to
6
M. Latifaj et al.
EcoreLanguage is equivalent to the EcoreEMF metamodel conforming to the Ecore meta-metamodel in EMF. By equivalent, we mean that the metamodels contain the same concepts and hierarchical structure, while the order by which metaconcepts persist in the metamodel file might differ. Step 2. Define automated mechanisms (i.e., model transformations) to transform EcoreMPS metamodels to Gen EcoreEMF metamodels and evaluate whether the latter is equivalent to the corresponding EcoreEMF metamodel and whether it can be correctly loaded and used in EMF. In addition, the generated Java classes from both metamodels (i.e., Gen EcoreEMF metamodel and EcoreEMF metamodel), should be the same. The implementation of EcoreLanguage in MPS and the transformations from EcoreMPS metamodels to EcoreEMF metamodels are described in Sects. 1.4.1 and 1.4.2, respectively, while the evaluation is described in Sect. 1.5. Fig. 1.2 Ecore language plugin in MPS
1.4.1
EcoreLanguage: Implementing Ecore in MPS
The first step towards providing a bridge between EMF and MPS is recreating the Ecore meta-metamodel as an EcoreLanguage in MPS. The structure of the EcoreLanguage consists of concepts, concept interfaces, and their corresponding children, properties, and references, as found in the Ecore meta-metamodel. It is noteworthy to mention that concept interfaces, similarly to interfaces in Java, cannot be instantiated in the EcoreMPS metamodel. Thus, the concept interfaces that need to be instantiated, are implemented as concepts. For each concept of the language, there is an editor that facilitates the manipulation of the abstract syntax tree (AST) and provides intuitive interaction. The constraints aspect is used to express advanced constraints that cannot be covered by the language structure. An example of that is the name of the metamodel. In Ecore, the name of a metamodel cannot contain characters such as underscores and white spaces. If they were used in the name of an EcoreMPS metamodel, the transformation would output an invalid Gen EcoreEMF metamodel, that would not correctly load in EMF. To mitigate this risk, we add constraints that limit the choice of valid characters in MPS. Moreover, to allow the initialisation of some properties/references/children to default values when a concept instance is created, we use concept constructors and rely on the behaviour language aspect in MPS. Upon complete implementation, the language is packaged as a plugin, that can be distributed to users. Fig. 1.2 illustrates how the implemented EcoreLanguage is included in the list of available languages to choose from when adding a new solution to an MPS project. By importing this language the user can start defining EcoreMPS metamodels.
1.4.2
Automatic Export from MPS to EMF
The implementation of the Ecore meta-meta model in MPS as EcoreLanguage is pivotal for cross-platform modelling across EMF and MPS. However, even though it allows the textual definition of EcoreMPS metamodels, the latter still remain “isolated”, as they cannot be used outside the MPS environment. To build a bridge and enable the exchange of metamodels between MPS and EMF, EcoreMPS metamodels are transformed to Gen EcoreEMF metamodels that can be correctly loaded and used in EMF. The transformations are defined in Java and driven by an implicit mapping that is used to define correspondences between elements of the source (i.e., EcoreLanguage) and target (i.e., Ecore meta-metamodel) languages. While defining these correspondences, it is important to fully understand the structure of both languages, thus, in the following we provide code excerpts from the definition of a metamodel, both in EMF and MPS. To simplify the reading, we describe the procedure by its instantiation on a specific example, the Family metamodel (depicted in Fig. 1.3 in terms of EcoreEMF). Listing 1.1 details the eClassifier Gender of type EEnum, whose values are restricted to eLiterals Male and Female from the family.ecore file. If we make a reference to Fig. 1.1, family.ecore represents the EcoreEMF.ecore artefact. Listing 1.2 details the Gender concept, from the familymodel. sandbox.mps file that represents the EcoreMPS. sandbox.mps artifact in Fig. 1.1, and it consists of two parts. 1. Language definition: Defines the languages used for the definition of EcoreMPS metamodels. For this specific
1 Cross-Platform Blended Modelling with JetBrains MPS and Eclipse Modeling Framework
7
Fig. 1.3 Family EcoreEMF metamodel
case, we have used a built-in language from MPS (line 3–5) that includes the BaseConcept and INamedConcept, and EcoreLanguage (line 6–11). EcoreLanguage reflects the Ecore meta-metamodel, where each concept can contain references, properties, and children, and they are all assigned randomly generated index values. Line 7 details the EEnum concept, while line 10 the EEnumLiteral concept. 2. Metamodel description: Defines the instances of concepts (i.e., nodes) that are described in the language. Each node stores a reference to its declaration, its concept. The node in line 15 stores a reference to EEnum, while the node in line 17, stores a reference to EEnumLiteral. To connect the language definition and metamodel definition, we use hash-tables as data structures that can map keys to values. Listing 1.1 XML Version of family.ecore 1 2
3
4
Listing 1.2 XML Version of familymodel.sandbox.mps 1 2 3
4 ... 5
6
7
8
9
10
11
12 13 14 15 16
17
18
19
20
21
22
23
24
25
26
8
Starting from the language definition, we iterate through all the elements of the language, and store the element’s index as key, and the element’s name as value. However, if we compare node and property/reference in the metamodel definition, we notice that they store references to their declarations differently. Nodes store references in the attribute concept, while properties and references store references in the attribute role. If we were to search for a node’s reference to its declaration in the role attribute instead of concept, that would raise an exception, because the role attribute does not exist. Thus, to mitigate this risk, we define two hashtables; one for conceptElements and one for propertyElements (used both for properties and references). Even in the event of changes to the EcoreLanguage, the implementation would detect the changes, and automatically update the hash-tables. The next step consists in defining a recursive function that leverages the tree-like structure of XMI files and traverses the nodes, starting from the root node in the XMI file and branching to the leaves. First, we access the get(Object key) method of the conceptElements hash-table, which returns the value to which the key is mapped in this hash-table. Recall that the key in the conceptElements hash-table is the index of the concept, while the value, is the name of the concept. Depending on the meta-object that equals the returned value, the implementation outputs an XML file that follows the same template as an EcoreEMF metamodel. Listing 1.3 details a code excerpt of the analyzeNode function. In lines 2 and 3 we have conditional statements that perform different computations, depending on the meta object that equals the returned value of the get() method. In the listing we only illustrate the “EPackage” and “EEnum”, but the complete implementation also includes “EEnumLiteral”, “EClass”, “EReference”, “EAttribute”, “EAnnotation” and “EOperation” meta objects. Lines 4-15 provide more indepth details regarding the computations that take place when the conditional statement that checks whether the returned value of the get() method equals “EEnum”, evaluates to true. The first step in this computation consists in creating an eClassifier and adding the attribute that identifies an EEnum in XMI to it. Next, we iterate through the childNodes of the node we are currently analyzing. If the childNode’s name is equal to “node”, the childNode is passed as a parameter to the analyzeNode function, and the appendChild() method is used to append this childNode to the list of children of the node under analysis. Else, we access the get() method of the propertyElements hashtable, where we pass as a parameter the value of the childNode’s role attribute. If the returned value equals to “name”, then we set the value of the childNode’s value attribute on the name attribute of the element.
M. Latifaj et al.
Listing 1.3 analyzeNode Function
1 public static Node analyzeNode(Node concept) { 2 if (conceptElements.get(concept. getAttributes().getNamedItem (“concept”).getNodeValue()).equals (“EPackage”)){...} 3 else if (conceptElements.get(concept. getAttributes(). getNamedItem(“concept”).getNodeValue()). equals(“EEnum”)) { 4 element = eclipseEcoreXML. createElement(“eClassifiers”); 5 element.setAttribute(“xsi:type”, “ecore:EEnum”); 6 for( int i=1;i 1h, in which vU is the meantime (in minutes) needed by the modelers to conclude instantiating the reference model; and the alternative hypothesis HA-1-1, represented by vU ≤ 1h.
61
One hour corresponds to six times the time spent by the researchers to instantiate, for the first time, the reference model based on the ACCU-CHEK Spirit. We also formalized the hypotheses to conduct statistical analyses related to the effort factor: the null hypothesis H0-1-2, i.e., vU > 3, in which vU represents the mean classification of the responses for the effort questions; and the alternative hypothesis HA-1-2, represented by vU ≤ 3. We also formalized hypotheses to conduct statistical analyses related to the RQ2. Considering understandability, there is a null hypothesis H0-2-1, i.e., vU ≤ 3, in which vU represents the mean classification of the responses for the understandability questions, while the alternative hypothesis HA-2-1 is represented by vU > 3. Considering adaptability, there is a null hypothesis H0-2-2, i.e., vU ≤ 3, in which vU represents the mean classification of the responses for the adaptability questions, while the alternative hypothesis HA-2-2 is represented by vU > 3. At the end of the experiments, we applied a questionnaire to collect metrics for the effort and reusability factors by providing a 5-point Likert scale (1) to (5). The scale interpretation concerning the metrics is based on the effort of instantiating the reference model during experiment 01: (1) represents the best result and (5) represents the worst result. For the understandability and adaptability measures, conducted during experiment 02, the scale’s interpretation is (1) for the worst and (5) for the best. The modelers answered the questionnaire after extending the reference model.
8.4.3
Analysis
Figure 8.2 depicts the answers for effort and each reusability factor evaluated in experiments 01 and 02, respectively. We did not apply more complex hypotheses tests such as the t-test and Wilcoxon test because the hypotheses were defined to evaluate RQ1 and RQ2 only use the mean responses concerning the factors considered in this experiment. The basic descriptive statistics enabled the evaluation of the hypotheses. Considering the RQ1, we evaluated the reference model’s reuse based on the analysis of the improvement of the productivity of the modelers. The 12 modelers who participated in the experiment concluded instantiating the model, and in the average case, within six minutes. When asked to respond to the questionnaire, 66.7% of the modelers (8 of the 12 modelers) stated that the task of instantiating the reference model presented no effort while 8.3% stated low effort, 16.7% stated medium effort, and 8.3% stated considerable effort. Therefore, the hypotheses H0-1-1 and H0-1-2 were refuted, showing that the reference model presented a positive response to the RQ1 (the modelers used only 10% of the estimated time). We evaluated the reference model’s reuse to analyze the RQ2 based on the understandability and adaptability factors. Figure 8.3 presents the number of modelers stating to extend
62
T. F. Costa et al.
Fig. 8.2 Answers distribution in accordance with the effort and reusability
Fig. 8.3 Instantiated model vs. correctly instantiated Fig. 8.4 Answers distribution considering complexity of functionalities
the model considering the requirements 01 and/or 02, also including the number of modelers who experimented with 02 correctly. All the modelers implemented requirement 01 (i.e., adding a new battery). However, only five of them implemented it successfully. The analyses of the work conducted by the seven modelers who did not implement requirement 01 successfully showed that they all failed in the same task. Six modelers attempted to implement requirement 02. However, only two of them achieved the correct answer. The results of the four modelers who did not achieve the correct implementation showed different mistakes, all related to the usage of CPN/ML. To analyze the time spent to conduct the second experiment, we considered the group related to each modeler. Only one of the modelers successfully implemented the reference
model’s first and second requirements within 25 and 38 minutes, respectively. The modelers scaled the first requirement as a low-complexity task while disagreed when asked about the complexity of the second requirement: the control group’s modeler scaled as a medium-complex task, while the modeler of the treatment group scaled it as a low-complex task. Analyzing understandability and adaptability, it is not possible to conclude that the reference model is fully reusable, requiring the analysis of results presented in Fig. 8.2. Most of the modelers did not fully understand each component of the reference model’s roles and functioning. The high number of modules required by the complexity of IIPS negatively impacted the task of adapting the model adding new requirements. Figure 8.4 supports this claim showing modelers’
8
A Model-Based Approach for Quality Assessment of Insulin Infusion Pump Systems
63
Fig. 8.5 GUI sample of the web-based application
answers for the complexity of each requirement added to the reference model in the second experiment. Most of the modelers stated the requirements 01 being of low (75%) or medium (25%) complexity, while stated medium or high complexity (75%) for requirement 02. The evaluation of RQ2 based on the questioning presented to the modelers showed that H0-2-1 was refuted due to the mean classification of the understandability factor (approximately 3.17). For H0-2-2, the mean classification of the adaptability factor presented approximately 3.58, refuting the null hypothesis. Thus, the reference model presented acceptable levels of understandability and adaptability.
8.5
Web-Based Application
Based on the results of the empirical evaluation, we used the ACCESS/CPN framework [11] and web services to develop a web-based application, aiming to assist modelers to re-
use our approach. The usage of ACCESS/CPN enabled us to embedded the reference models in web services to conduct simulations without the GUI of the CPN/Tools software. Figure 8.5 presents a GUI sample of the web-based application. For instance, the application enables modelers to simulate the configuration of the pump, showing each simulation step (e.g., along with information related to places and transitions). The web services used to implement such application can be easily adapted to handle a different version of the reference model. The modeler can use the system to handle the CPN model of IIPS using four features: (1) add battery; (2) remove the battery; (3) configure the insulin pump; and (4) simulate the insulin pump. Features (1) and (2) allow the user to change the structure of the CPN model by adding/removing an instance of a module that represents the battery of the insulin infusion pump. Feature (3) allows modelers to change the values of the parameters of the reference model. Finally, Feature (4) allows modelers to simulate the CPN model using the GUI, without using CPN/Tools, to improve
64
T. F. Costa et al.
the understanding of the pump’s behavior without the user having access to the internal structure of the model.
8.6
Conclusions and Future Work
We presented an MBA focused on CPN reference models of IIPS, aiming to assist manufacturers in assessing quality. The case study was relevant to extending the model to verify and validate two model refinements as quality assessment scenarios. We considered the formal verification of two safety properties for the first refinement using the model checking technique. We validated the second refinement using simulations to analyze the model concerning the commercial system’s technical specifications. Finally, the empirical evaluation demonstrated that the MBA/CP N is reusable for the development and certification process of IIPS. As future work, we envision conducting another empirical study to evaluate the web-based application, integrated with the reference model, to analyze if understandability and adaptability are improved. Acknowledgment This work was supported in part the CNPq and CAPES.
References 1. V.R. Basili, H.D. Rombach, The TAME project: towards improvement-oriented software environments. IEEE Trans. Softw. Eng. 146, 758–773 (1988) 2. X. Gao, Q. Wen, X. Duan, W. Jin, X. Tang, L. Zhong, S. Xia, H. Feng, D. Zhong, A hazard analysis of class I recalls of infusion pumps. JMIR Hum. Factors 62, 10366 (2019) 3. Infusion Pump Improvement Initiative. https://bit.ly/2QAHe2V 4. Infusion Pumps Total Product Life Cycle. https://bit.ly/2EMOXrW 5. K. Jensen, L.M. Kristensen, Colored Petri nets: a graphical language for formal modeling and validation of concurrent systems. Commun. ACM 58(6), 61–70 (2015) 6. B. Kitchenham, L. Pickard, S.L. Pfleeger, Case studies for method and tool evaluation. IEEE Softw. 124, 52–62 (1995) 7. MBA/CPN Repository. https://bit.ly/2Qyci3k 8. Pump User Guide - Accu-Chek. https://bit.ly/2QzJMOT 9. The Assurance Case Working Group, Goal Structuring Notation Community Standard (Version 2) (2018) 10. H. Washizaki, H. Yamamoto, Y. Fukazawa, A metrics suite for measuring reusability of software components, in Proceedings of the 5th International Workshop on Enterprise Networking and Computing in Healthcare Industry, Sydney, Australia (2003), pp. 211–223 11. M. Westergaard, L.M. Kristensen, The access/CPN framework: a tool for interacting with the CPN tools simulator, in Proceedings of the International Conference on Applications and Theory of Petri Nets (2009)
9
Narrowing the Gap Between Software Engineering Teaching and Corporate Environment Marcelo A. M. da Conceicao, Oswaldo S. C. Neto, Andre B. Baccarin, Luan H. S. Dantas, Joao P. S. Mendes, Vinicius P. Lippi, Gildarcio S. Gonçalves, Adilson M. Da Cunha, Luiz A. Vieira Dias, Johnny C. Marques, and Paulo M. Tasinaffo
Abstract
Currently, there is a gap between university education in the area of software engineering and the corporate and industrial environment. In this way, it is necessary to explore new learning models to transmit complex knowledge more quickly and interconnect these different environments. For that, the STEPES-BD was developed on the first half of 2020, in the midst of the COVID19 Pandemic. In this project, the students have put into practice some concepts from three different undergrad and graduate courses of Computer and Electronic Engineering in the Area of Informatics of the Aeronautics Institute of Technology (Instituto Tecnológico de Aeronáutica – ITA) in Brazil. To approach these different worlds, it was used the Interdisciplinary Problem-Based Learning (IPBL), the Scrum Framework (SF), and also several emerging Information Technologies of public domain. This article aims to present the main results obtained by students in just one semester of 17 weeks, in a collaborative and cooperative manner, allowing them to implement a computer system prototype to help solve a real-world problem of managing health information for the COVID-19 Pandemic. In this article, it is described how the IPBL, the SF, its ceremonies, roles, and artifacts were adapted. The assessment of students experience in the STEPES-BD Project was carried out using the Tuckman Model, extracting quantitative and qualitative M. A. M. da Conceicao () · O. S. C. Neto · A. B. Baccarin L. H. S. Dantas · V. P. Lippi · G. S. Gonçalves · A. M. Da Cunha L. A. Vieira Dias · J. C. Marques · P. M. Tasinaffo Computer Science Division, Aeronautics Institute of Technology – ITA, São José dos Campos, Brazil e-mail: [email protected] J. P. S. Mendes State Technology Educational Center Paula Souza, São José dos Campos, Brazil
results, comparing the results achieved with related works, and acquiring students’ perceptions through project-end questionnaires. Keywords
Software engineering education · Computer science education · Interdisciplinary problem-based learning · Scrum · Agile · Soft skills · Tuckman model maturity level · Scrum of scrum · Storytelling · Emerging technologies
9.1
Introduction
Rapid technological and social changes have demanded constant updates in curricula of universities and research institutes, causing the need for new educational configurations. These demands have forced who is involved in these processes to explore new learning models to transmit complex knowledge more quickly and to interconnect the academic world with the real world of companies, public, or private institutions. The Interdisciplinary Problem-Based Learning (IPBL) approach has been providing these interconnections, through student-centered learning processes, with emphasis on different problem situations [1]. This approach has reflected the complex reality faced by undergraduate and graduate students¸ allowing more realistic contextualization and more intense collaborations between students to apply, even in their academic environments, concepts obtained from disciplines into real-world problems [1]. Within this context, this article describes how the IPBL and the Scrum Framework (SF) were adapted and used in just one academic semester of 17 weeks for developing a computer system prototype involving a health problem of infor-
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_9
65
66
M. A. M. da Conceicao et al.
mation management, by using some emergingtechnologies of public domain as Blockchain among others in the scenario of the COVID-19 Pandemic. For this, the STEPES-BD was conceived as an academic project using Specific Technological Solutions for Special Patients and Database Systems. This interdisciplinary project brought together 25 students (84% male and 16% female) from the following undergrad (12%) and graduate (88%) courses of Computer and Electronic Engineering in the Area of Informatics at the Brazilian Aeronautics Institute of Technology: CE-245 Information Technologies; CE-240 Database Systems Project; and CE229 Software Testing. Working collaborative and cooperatively, these students were able to develop a product as a proof of concept at the level of system of systems. During the development of this project, the maturity level reached by each interdisciplinary Scrum Team (ST) was evaluated based upon the Tuckman Model Maturity Level (TMML) and artifacts, roles, and ceremonies were adapted and audited by the professors so that the development process could be measured in the most possible practical and natural manner. In order to contextualize and guide the academic project, the following scenario was described as an Assigned Mission storytelling: “For public or private organizations involved with the management of information of interest to Patients, Physicians, Hospitals, and Suppliers, which provide monitoring and decisions in the area of health knowledge, the STEPES-BD involves a computational system based on emerging technologies. Unlike other products that may exist in Universities, Research Institutes, Government Agencies, or private companies, this product was developed and tested in 17 academic weeks of the 1st Semester of 2020.
9.2
Related Work
Francese et al. [2] report on the experience acquired in the semi-attendance Mobile Application Development Course for 55 students in 12 weeks, using PBL learning. The students were divided into 27 teams in pairs form, creating intrateam collaborations and an extra-team competitions for the best application. Teams’ evaluation was performed by the professors, by the juries composed of 10 IT professional managers of national and international companies, and by the students through answers to a project-end questionnaires. Kizaki et al. [3] report collaborative software development through PBL learning to undergraduate students, where 32 students participated and were divided into groups from 3 to 5 students, having 1 Project Manager and 1 Scrum Master per each team, as a member-of-society engineer. The evaluation of projects was performed according to the Kano Model [4]. Raibulet [5] describes the experience of collaborative activities with students of the Software Engineering course,
aiming to limit the gap between industry expectations and the preparation of undergraduate students. The project lasted 10 weeks, where 41 students were divided into groups from 3 to 5. Each team has developed its own web software project. The project evaluation was performed through questionnaires, capturing feedback from students’ perception.
9.3
Tailoring Scrum and IPBL
In order to overcome the needs imposed by the social distancing due the COVID-19 Pandemic, in this project the Totally Integrated Scrum model was addressed, in which ST are multi-functional, having members distributed geographically, requiring strong extra-teams communications, as well as the adoption of Scrum of Scrums (SoS) meetings [6]. To support the interaction between the students, two levels of Scrum were created: (SoS Level #1) traditional intrateams; and (SoS Level #2) extra-teams, involving only the Product Owner (PO) and the Teams Scrum Masters (TSMs). For the extra-team integration and for the geographically dispersed members to be integrated, SoS meetings [7] were held where, in large projects with multiple teams like the STEPES-BD, you may create a higher level of aggregation between the PO and the TSMs, as shown in Fig. 9.1. Furthermore, the STEPES-BD Project has applied the IPBL pedagogical concepts as its basis, integrating students from the three different disciplines, using distinct and emerging technologies for the development of a Proof of Concept (PoC). In this project, it was addressed the real-world problem of the COVID-19 Pandemic that has been afflicting the populations of the world in recent years. As it is an academic project with a well-defined restriction in its schedule and requiring reductions in scope, its management was always carried out by using the Scrum best practices. That’s why the time boxes between the ceremonies and the artifacts had to be adapted to allow students a greater opportunity to get in touch with the Scrum Agile method. For this, the STEPES-BD Project was divided into four ST representing the interest segments in the health area: Patient, Physician, Hospital, and Supplier. Interdisciplinary Activities Although STs had the specific scopes of their well-defined segments, they had to work together, starting from the 2nd Sprint, performing interdisciplinary activities, mainly to properly address common efforts, involving integration. Team Arrangement The choice and distribution of students to join STs were carried out in a balanced way by professors who were in charge of balancing and adjusting any differences in skills and experiences existing within STs.
9 Narrowing the Gap Between Software Engineering Teaching and Corporate Environment
67
Student's Maturity Level Assessment (According to the Tuckman Model)
Inception
New Academic Semester
Product Owner
choose
SoS Level #2 Teams' Scrum Master
Real-World Problem
Product SoS Level #1
Final Presentation Weekly Meetings
Storytelling
60' minutes
Teams
guide
Team Arrangement
Skill Leveling
Sprints
Sprint 0
guide
Sprint Planning Meetings
Burndown Chart
Individual Syntetic Report
Sprint Backlog
Scrum Team Syntetic Report
Sprint Review
Kanban
Week #1
Sprint Review Sprint Presentation Retrospective Sprint Review Videos
...
Week #4
Sprint 1 - 3
Strengths Weakness Suggestions
4-Week / Sprint
Product Backlog
Project Review Meeting Final Individual Report
Videos
Final
Fig. 9.1 The steps adopted in STEPES-BD project
(A) The Scrum Roles Academically Used To enable the application of Scrum in the academic environment, 2 new backup roles had to be created (Backup of Product Owner and Backup of Team Scrum Master), to provide eventual replacements of STs key members. The selection and choice of students to play the roles of PO, PO Bkp, the four roles of TSMs, and the four roles of TSMs Bkp was carried out by professors of the disciplines based upon skills and qualifications proven in their curricula and in previous participation experiences in the development of Agile Projects, using similar structures as those quoted in [8,9]. During the STEPES-BD Project development, professors acted interested or directly involved with the Project (as Stakeholders) and constantly monitored the progress of their development, through the artifacts generated at each Sprint, as Burndown Charts, Sprint Backlogs, and Test Case Spreadsheets.
Sprint Backlogs It was generated from the Sprint Planning Meetings of each ST, containing USs selected to be developed during the Sprints. They typically involved information on USs development priorities, from scores obtained in effort planning estimates to be spent, such as Planning Poker, Integrations, USs Definition of Ready, USs Definition of Done, Acceptance Criteria, and so on. Planning Poker Games It was generated in the Sprint Plannings Ceremonies, within each ST and after the US was subdivided into Tasks. In them, the Planning Poker technique was applied, where the scores, consensually and according to the Fibonacci Sequence, were assigned, representing the necessary efforts for the complete development of each US. Burndown Charts It was generated in the Sprint Plannings Ceremonies, within each ST and after the US was subdivided into Tasks. It allowing Stakeholders, POs, and team members to visually monitor the effort spent, inferring the existence of delays or advances in the scores made from USs.
(B) The Scrum Artifacts Academically Used The following artifacts, sorted chronologically, have also been used and adapted to meet the academic needs of the STEPES-BD Project, as shown in Fig. 9.1. Product Backlog It was created in the beginning, updated during the development of the project, containing a dynamic and updated list of User Stories (US) that could be implemented in the project, as well as information about possible inter-teams integrations, acceptance criteria, and dependencies.
Kanban It was generated and updated during Sprints. In this example, all USs and Tasks were registered on GitHub, right after a Sprint Planning Meeting. In this example, it is noticed that the ST used this Kanban board to hold their Weekly Meetings, indicating those responsible for their developments, and adjusting their progress as To Do, In Progress, and Done. Individual Synthetic Reports It was prepared, updated, and presented individually by the Team Members at the end of each Sprint and during the Sprint Reviews, containing a summary of what was developed during the Sprint and individual suggestions for improving the Project.
68
Scrum Teams Synthetic Report It was prepared, updated, and reported by the TSMs, at the end of each Sprint, containing the information consolidated by Teams, in a similar manner to the Individual Synthetic Reports previously described. Sprint Review Presentations It was prepared by the Team Members and presented by the TSMs, at the end of each Sprint and during the Sprint Reviews. They were created to report developments made by the ST during Sprints. Sprint Review Videos It was prepared by the Team Members and also presented by the TSMs, at the end of each Sprint and during the Sprint Reviews. They were presented in 3-min videos, containing a summary of what was developed. Video It was presented in an integrated manner by the TSMs, in the Final Presentation of the Project, in the form of a 5-min video, consolidating all the value deliveries. Final Presentation It was created and presented in a collaborative manner by the TSMs, in a Final Presentation, lasting approximately 60 min, where all value deliveries were consolidated. Final Individual Report These last artifacts were created and published individually by the students, containing: a summary of all the development carried out; the student participation in each US; and its conclusions and suggestions for the next projects.
M. A. M. da Conceicao et al.
Burndown Charts was presented, impediments were checked, and so on. At the end of the Sprints, in the last weeks, the STs carried out the Sprint Reviews, where it was possible to show the USs developed and the main artifacts such as Burndown Charts. Each ST developed a video, in addition to their technical reports, summarizing the following topics: (1) sprint’s contextualization and objectives; (2) Sprint Backlog; (3) requirements; (4) US and associated Tasks; and (5) suggestions. In the Sprint Retrospective were carried out through Google Forms and their results presented, containing strengths and weaknesses, in addition to suggestions for changes for the following Sprints and future works. The Project Review Meeting, the final results obtained were presented. To overcome communication challenges in a completely distributed environment, the team members made massive use of online communication tools. (D) The STEPES-BD Scrum Ceremonies To assess the maturity levels achieved by the STs, considering the participation of its members in the ceremonies and in the uses of the artifacts, the TMML was applied. When applying this model, it was found that, although the STs sought to practice self-management, there was a need to verify the maturity levels achieved by each of them, in order to carry out behavioral realignments and evaluations [10, 11].
(C) The STEPES-BD Scrum Ceremonies To manage the work carried out during the Sprints, weekly, on Mondays, remotely and through video-conference tools, the PO and the TSMs held the Weekly Meetings (WM) on the second level of SoS (SoS Level #2), to present the project progress, as shown in Figure 1. At these meetings, in addition to discussing relevant topics to all ST, general information and guidelines were regularly received from professors. At the first meeting of each Sprint (into SoS Level #2), US were selected to compose the Sprint Backlog for each ST. The intra-team meetings (SoS Level #1) were held weekly, one day after the extra-team Weekly Meetings (SoS Level #2), replacing the traditional Daily Meeting, both by video-conference manner. TSMs and their developers participated in this ceremony. In the first meetings of each Sprint, called \textbf{Sprint Planning Meetings}, ST played Planning Poker, created Burndown Charts, Kanban boards were made available, and the STs Sprint Backlogs were prepared with the additional information generated in the meetings. In the others Weekly Meetings of Sprints intra-teams, in level #1 of SoS, the STs progress and Kanban was updated,
Forming Already on the first Sprint, after explaining the ST objectives within the project, one can verify manifestations within the teams of the main characteristics of this stage. In this stage, team members usually start their participation with fear of how to adapt to the environment. To overcome this difficulty, acting as facilitators, professors, PO, and TSMs needed to support their members by clarifying doubts and providing technologies that could be initially used to facilitate the tasks to be performed. Storming In any project that involves group dynamics, conflicts can arise within the teams. In the TMML, these conflicts characterize the second stage of the maturity level where each member knows the goals for that moment, but cannot clearly define their individual role or responsibility, generating conflicts that have to be resolved by leaders within the groups. In this Project, evidence of this stage only began to be noticed at the end of the first and the beginning of the second Sprint and it was, once again, the responsibility of professors, the PO, and the TSMs to support their members, to clarify doubts and provide technologies, always as facilitators of the tasks to be performed.
9 Narrowing the Gap Between Software Engineering Teaching and Corporate Environment
Norming In this third stage the Team Members already know each other and began to deal better with possible situations that could occur. In this stage, the processes, and execution of tasks became more natural. The members were already more motivated by the first results achieved. At this point, the leaders needed to actuate when necessary. In this Project, evidence of this stage only began to be noticed at the end of the second and the beginning of the third Sprint, with an ever smaller participation of the professors, the PO, and TSMs as facilitators in the project. Performing The fourth stage is characterized by the full and self-sufficient performance of all members of a team, when they are able to autonomously develop their roles and manage themselves without the the need of leaders to intervene, because each participant has a well-defined role, knows the goals and objectives to be accomplished, knows how to work collaborative and cooperatively, and understands the importance of their participation to successfully deliver values. Some evidence of this stage only began to be noticed at the end of the third Sprint, but just within two STs. Adjourning The fifth and final stage consists of the dissolution of a group due to greater force, withdrawal or closure of a project. In this project, this stage occurred naturally, at the end of the project and at the end of the academic semester, giving rise to the beginning of the preparation of this article by those interested in documenting for the academic and scientific community the experiences acquired.
9.4
Proof of Concept
This section describes the features of the STEPES-BD Project, such as periodic deliveries developed over four Sprints. The Sprint Zero This Sprint was dedicated to leveling the knowledge of the students, designating them to: carry out short courses through the Internet; identify and develop skills in Team Members; receive the initial guidelines for the formation of teams and definition of the Scrum roles. In this Sprint, the initial versions of the Assigned Mission artifact was defined to make a single delivery of value at the end of the Project. In addition, the following were also elaborated: Product Backlog, involving US that could turn into future functionalities of a computational system, as well as the Project Architecture artifact. At the end of this Sprint, some students with interpersonal characteristics and appropriate techniques to be allocated to the teams and segments of the Project have already started to stand out.
69
The Sprint The teams chose to be conservative, which was corroborated by the only 157 points selected for the 13 USs, during the first Sprint Planning, applying the Planning Poker technique. These USs were divided into 74 tasks. In this Sprint, according to the TMML, the ST have already managed to migrate from the Forming stage to the Storming stage. The latter, characterized by conflicts of ideas and working methods among its members [10]. For a better use of students’ learning, this Sprint was chosen so as not to make extra-team integrations yet. In this way, each ST started to develop its own front ends, data models, back ends, test cases, and web servers to support the computational system. Also, each ST individually developed their functionalities to enable surveys and management from your registration data. The Sprint 2 In this Sprint, after applying the Planning Poker technique, the STs estimated a total of 226 points for the 14 USs selected in the Sprint Backlog #2, which enabled the development of 101 Tasks, representing an increase of 44% in implemented points in relation to the previous sprint. In this Sprint, it was found that members of the STs were already more integrated and able to work better as teams, starting the integration between teams by pairs of segments (Patient with Physician and Hospital with Supplier). However, teams still have needed some interventions from their TSMs to better standardize and organize their meetings and activities between teams showing that they were still in the Storming stage and only walking towards the Norming stage. Considering the TMML, the STs achieved their goals and managed to migrate from Sprint 1 and from the Forming stage to the Storming stage. Due to this good performance in Sprint 1, it is believed that, perhaps due to overconfidence, the STs have decided to increase and overestimate (by another 44%, in relation to the previous Sprint) the total points of their estimated efforts to be performed in this Sprint 2. For these reasons, it is believed that the STs have not been able to achieve a better performance nor have they migrated yet to the Norming stage in this Sprint 2. The Sprint 3 In this Sprint, after applying the Planning Poker technique, the ST estimated a total of 256 points for 17 US selected in the Sprint Backlog #3, which led to the development of 129 Tasks, consolidating an increase of 13% in the points implemented in relation to Sprint 2 (previous). It was verified that all STs were already in Stage 3, Norming, and out of that only 2 of the 4 STs had some signs and characteristics inherent to the Stage 4, Performing, as they already interacted with each other without the need for intermediation by the PO, professors or the TSMs. This
70
M. A. M. da Conceicao et al.
behavior can be seen from various alignment and integration meetings recorded in the Minutes, where members of some STs already called spontaneously and when necessary. In this Sprint, deliveries with greater aggregate value could be observed. This was mainly due to the requirements, chosen in terms of USs to promote the development of more consolidated Tasks also aligned with the maturity levels reached. Specifically with regard to the relationship between the teams and differently from what happened in Sprint 2, only with integrations in pairs (Patient with Physician and Hospital with Supplier), in this Sprint 3, integrations started to occur between all STs simultaneously, characterizing interoperability. Finally, another relevant factor for this project to be successfully completed was the situational awareness and the maturity levels progressively achieved by the STs over the 3 Sprints. In planning for an increase of just 13% of points in Planning Poker on Sprint 3, compared to Sprint 2, the STs have shown better accuracy over time to correctly estimate the efforts to be made for the development of USs and Tasks. The Final Presentation It took place on June 30, 2020, in the form of a video-conference and with the live participation of 40 people, including the Rector of the Aeronautics Institute of Technology, students, professors, and Higher Education Institutions in the Paraíba Valley region, São Paulo, Brazil, among other guests interested. A 60-min recording of the Final Project, where nine students representing the others have linked the main topics of this Project. The end of the project was marked by the dissolution of the teams, Adjourning stage [11], due to the successful conclusion of the academic semester.
9.5
Results and Discussion
This section presents the outcomes gathered from different perspectives: (1) quantitative results achieved from measurements of worked hours, completed tasks, and accomplished points; (2) students’ feedback synthesis to understand the evolution of the students at the end of the project; and (3) comparison with the related works shown in Sect. 9.2. (A) Quantitative Outcomes In order to generate the quantitative outcomes, data from the ST #2 (Physician) were synthesized, because have presented similar performance to the other teams, in addition to better detailing of worked hours in each US. It was possible to analyze that students in the 3rd Sprint needed fewer hours to complete the tasks. The speeds were 4.1, 3.3, and 2.7 h per point (hpp), in Sprints 1, 2, and 3,
respectively. This represented an increase in speed of 20% and 34% on Sprints 2 and 3, respectively, as compared to Sprint 1. This speed improvement was corroborated by the evolution of the teams, in which the teams became more mature and improved their management. (B) Students’ Feedback In order to collect feedbacks from the students’ perception, at the end of the project, students answered questionnaires in which the responses were synthesized into the following questions categories: C1: What worked well? and C2: What can be improved? In Category 1 (C1), the responses were grouped into 7 topics and 43 mentions were collected: extra-teams integration (33%), union (21%), maturity (16%), meetings quality (9%), communication (9%), organization (7%), and engagement (5%). In Category 2 (C2) the responses were grouped into 10 topics and 39 mentions were collected: individual assessment (38%), communication (23%), overload (15%), and artifacts enhancement and technology specification (5% each). The other topics received only one mention each. From these results, it can be analyzed that the main objective of the STEPES-BD Project was achieved, as it is precisely in soft skills that the main gap between academic experience and the development environment in companies and industries are. The STEPES-BD Project achieved a good rate of students’ perception in the topics of extra-team integration, union, and maturity. Otherwise, there are some improvements to do, as pointed by the students, especially in the manner of the evaluation was conducted, where the individual assessment was the main suggestion. In addition, communication was mentioned, indicating that communication processes need to be adjusted. Finally, the topic of overload was mentioned a lot. This is mainly due to the short project time and the learning curve of some technologies. (C) Related Work Comparision The Table 9.1 describes a comparison between the related works and the STEPES-BD Project, represented by the last record. The Y column displays the year of the article. The Cit. column displays the number of citations based on Google Scholar. The Part. column shows the number of students participating in the projects. The Teams column indicates the number of teams participating in the projects. The Dur. column shows the project duration, where “y” and “w” represent years and weeks, respectively. The Type column indicates whether the approach was composed of a single collaborative project (S) for all students or multiple distinct projects, one for each team (M). The Mode column shows whether the project was executed in-person (P), semi-attendance (S) or remotely (R). The Eval. column indicates how the projects
9 Narrowing the Gap Between Software Engineering Teaching and Corporate Environment Table 9.1 Quantitative comparision with the related works # [3] [2] [5] S
Y 14’ 15’ 18’ 20’
Cit. 16 48 20 –
Part. 32 55 41 25
Teams 7 27 8 4
Dur. 1y 12w 10w 17w
Type M M M S
Mode P S P R
Eval. K E, F, P F, P F, P
were evaluated, where the values indicate: “E” – external members assessment; “F” – qualitative assessment by students’ feedback through questionnaires; “K” – Kano Model; “P” – artifacts evaluation by professors. By analyzing Table 9.1, one can see the different strategies used by the 4 works. Regarding the number of students in the project, it can be seen that the project by Francese et al. [2] was broader (55 students), while in this Project had 25 students. 75% of works were carried out in just 1 academic semester: STEPES-BD Project, [2], and [5]. Only in Kizaki et al. [3], the project was longer than 1 semester and 75% of the work have used different projects for each team, differing from the STEPES-BD Project, in which all students worked cooperative and collaboratively in the same project. This can be observed by the reduced quantity of the STEPES-BD Project teams (4), against 27, 7, and 8 of other works, [2, 3, 5], respectively. Regarding the students’ interaction manner, an exclusively remote modality was adopted, due to the COVID-19 Pandemic. In other works, the in-person modality (50%) [3, 5], and semi-attendance (25%) [2] were used. For projects’ evaluation, feedback from students through questionnaires and artifacts assessment by teachers (75%) were reported as the main form of evaluation: STEPES-BD Project, [2], and [5]. Francese et al. [2] also added the evaluation by external collaborating participants. Only in Kizaki et al. [3], the evaluations of others were not used, because they followed the Kano Model. The originality offered in the STEPES-BD can be seen as students were harmoniously integrated into a single project in interdisciplinary teams. In addition, carrying out the project in a completely remote environment may reflect the challenges of society today, in which the home office has been expanding. Thus, the real applicability of the approach used in the STEPES-BD Project for teaching Software Engineering disciplines is evident, providing an approximation of teaching in universities and the real needs of companies and industries.
9.6
Conclusion and Future Work
This article described the development of the academic project Specific Technological Solutions for Special Patients and Database Systems, named in Portuguese STEPES-BD
71
Project, from an Interdisciplinary Problem-Based Learning approach (IPBL). In it, the best practices of the Scrum management method were used, providing undergrad and graduate students at the Brazilian Aeronautics Institute of Technology (ITA) with a practical academic experience involving the development of computational systems, using emerging technologies. At the end of the semester, the students managed to demonstrate with their Scrum Teams the successful development of a relatively complex computational system, in terms of a Proof of Concept, in a self-managed manner. To enable the development of this Project, the Scrum framework was adapted to support integrations and interdisciplinarity in 2 levels of Scrum of Scrum, providing students an opportunity to participate in an agile project, involving Information Technologies to meet the needs of society, and public or private companies. For the natural continuity of this Academic Research and Development Project, it is recommended to investigate other metrics for performance evaluation to compare results to be achieved with the good practices of the Scrum method. As future work, it is suggested that IPBL applications be further explored and deepened in other ITA disciplines of Undergrad and Graduate Programs. Acknowledgments The authors of this article would like to thank the Brazilian Air Force, the Brazilian Aeronautics Institute of Technology, the Ecossistema Negocios Digitais Ltda (WiBOO/Wibx) and the Casimiro Montenegro Filho Foundation, for the support and infrastructure offered during the development of this ALFA Project (MEC-ITA).
References 1. M. Brassler, J. Dettmers, How to enhance interdisciplinary competence – Interdisciplinary problem-based learning versus interdisciplinary project-based learning. Interdiscip. J. Prob. Based Learn. 11, 9 (2017) 2. R. Francese, C. Gravino, M. Risi, G. Scanniello, G. Tortora, Using project-based-learning in a mobile application development course – An experience report. J. Vis. Lang. Comput. 31, 196–205 (2015) 3. S. Kizaki, Y. Tahara, A. Ohsuga, Software development pbl focusing on communication using scrum, in 2014 IIAI 3rd International Conference on Advanced Applied Informatics (2014), pp. 662–669 4. N. Kano, Attractive quality and must-be quality. Hinshitsu (Qual. J. Jpn. Soc. Qual. Control) 14, 39–48 (1984) 5. C. Raibulet, F.A. Fontana, Collaborative and teamwork software development in an undergraduate software engineering course. J. Syst. Softw. 144, 409–422 (2018) 6. J. Sutherland, A. Viktorov, J. Blount, N. Puntikov, Distributed scrum: Agile project management with outsourced development teams, in 40th annual Hawaii International Conference on System Sciences (HICSS’07) (IEEE, 2007), pp. 274a–274a 7. J. Sutherland, Inventing and reinventing scrum in five companies. CutterIT J. 14, 5–11 (2001) 8. G.S. Goncalves, R. Shigemura, P.D. da Silva, R. Santana, E. Silva, A. Dakwat, F. Miguel, P.M. Tasinaffo, A.M. da Cunha, L.A.V. Dias,
72 An agile developed interdisciplinary approach for safety-critical embedded system, in Information Technology-New Generations (Springer, 2018), pp. 947–949 9. L.S. Siles, M.V. Santos, R.A. Rodrigues, A. Lineu Filho, J.P. Siles, R.E. Maria, J.C. Marques, L.A. Dias, A.M. da Cunha, An integrated academic system prototype using accidents and crises managementas pbl, in Information Technology-New Generations (Springer, 2018), pp. 419–427
M. A. M. da Conceicao et al. 10. B.W. Tuckman, Developmental sequence in small groups. Psychol. Bull. 63(6), 384 (1965) 11. B.W. Tuckman, M.A.C. Jensen, Stages of small-group development revisited. Group Org. Stud. 2(4), 419–427 (1977)
API-First Design: A Survey of the State of Academia and Industry
10
Nicole Beaulieu, Sergiu M. Dascalu, and Emily Hand
Abstract
Keywords
The evolution of distributed and cloud-based systems has led the computing community to converge on Microservice Architecture (MSA) as a preferred solution to distributed software design. Established design methodologies applied to MSA (e.g., Data-, Model-, and DomainDriven Design) assist in decision-making about business capacity and functionality encapsulated by the microservice. An expected result of microservice design is a welldefined Application Programming Interface (API) that facilitates the access of microservice and system capabilities. However, even with the extensive documentation and defined frameworks guiding practitioners in their execution of MSA, challenges exist in defining and exposing clean APIs. Further, the industry’s current focus on maximizing business capacities exposed by distributed systems emphasizes the importance of improving API design and implementation. To this end, API-First Design is emerging as a viable approach to MSA and API design. API-First principles suggest that all capabilities of an organization and its systems are exposed via an API and that the foundation of system design is the definition of clear and welldefined APIs. A significant challenge associated with API-First Design lies in the infancy of the topic and the necessity for peer-reviewed research defining guidelines for adoption and a baseline for future research. This paper seeks to move the state of the API-First Design methodology forward by exploring publications of the academic community and grey literature available on the topic. The paper concludes with a discussion about future research opportunities that may advance the understanding and adoption of API-First Design.
API-first design · API-driven design · Software architecture · Software design · Software engineering · Microarchitecture · Microservice · Microservice architecture · API culture · API design
N. Beaulieu () · S. M. Dascalu · E. Hand College of Engineering, University of Nevada, Reno, Reno, NV, USA e-mail: [email protected]; [email protected]; [email protected]
10.1
Introduction
Over the last few decades, as systems have become increasingly distributed and data-intensive, Service Oriented Architecture (SOA) has become a standard for solving the technical challenges of industry enterprise software. Today, enterprise systems encapsulate business functionality, rules, and domain logic in standalone, loosely coupled, and independently deployable applications known as microservices [1–3]. As the popularity of microservices has increased in distributed system design, their benefits, limitations, and applicability have caught the attention of academia and are a topic of modern research. Studies include design methodologies that assist practitioners and researchers alike in making savvy decisions about the functionality of and boundaries between microservices that comprise a system. Such methodologies include Model-Driven [4, 5] and Domain-Driven [6] Design. Each methodology defines a set of practices intended to maximize the benefits of microservice architecture; the approaches focus on the resulting services and their encapsulated business capacities [2, 3, 7, 8]. However, attention to the interfaces (APIs) through which services interact, are accessed, and their capacities exposed for the business is often an afterthought [1, 3]. Added is the complexity of modern distributed systems, their APIs, and the variety and scale of target device requirements as shown in Fig. 10.1.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_10
73
74
N. Beaulieu et al.
Fig. 10.1 The modern API
An emerging methodology, API-First Design, is gaining the attention of software practitioners in the industry. Its principles suggest that all business capabilities supported by a system are exposed via an API. The methodology also suggests that design focuses on the definition of clear and well-defined APIs as its foundation [7, 9]. A challenge associated with API-First Design lies in the imbalance between industry-generated grey literature (documentation that has not undergone the rigor of academic peer review or is not included in the scholarly publications traditional in academia [10]), and peer-reviewed research that may provide guidelines for adoption and a baseline for future research. Thus, given the importance of APIs in modern software solutions and the focus on API-First Design as a viable methodology for SOA and MSA, a systematic review of grey literature provides valuable context and content. For career software professionals and practitioners of software engineering, the exploration of API-First Design presents an opportunity to explore end-to-end the possibilities created by this relatively young design methodology. This paper aims to review the existing academic research and grey literature, provide a comprehensive set of references, and identify options for potential future research of API-First Design.
This publication, in its remaining sections, is organized as follows: Section II provides a background on software architecture and methodologies; Section III overviews the state-of-the-art of API-First Design in academia and industry grey literature; Section IV outlines potential directions of future research; and Section V closes with a summary of the paper’s main findings and our concluding remarks.
10.2
Background
The concept of Software Architecture emerged in the 1990s [11]. Its introduction provided lexicon and methodologies for describing software and its complexities from a high level. The concept of Service-Oriented Architecture (SOA) was introduced in 1996; its tenets include the organization and design of business functionality for distributed systems [11,12] as well as the stability of the service Application Programming Interface (API) [11]. With the ubiquity of cloud-based, distributed, and data-intensive systems, SOA has become a standard for industry enterprise software design [1]. Specifically, systems frequently encapsulate business functionality, rules, and domain logic in loosely coupled and independently deployable applications known as microservices. A succes-
10
API-First Design: A Survey of the State of Academia and Industry
75
sor to SOA, Microservice Architecture (MSA) is a software architecture pattern introduced and coined in 2014 by Martin Fowler [1]. MSA more concisely defines a distributed architecture than the monolithic implementations of the SOA-era distributed enterprise systems. Microservice goals include encapsulated business capabilities, loose coupling between components, lightweight communication mechanisms, endpoint intelligence, and deployment independence [1]. Many modern enterprise software systems take advantage of the benefits of MSA, including those from Netflix, Amazon, Guardian, Twitter, [2, 3], Intuit, and Figure Technologies.1 The practices and methodologies for designing microservices are well-researched and documented. Common design approaches include Model-Driven Design (MDD) [5] and Domain-Driven Design (DDD) [6]. Introduced in approximately 2001 and an extension to MDD, MDA®[4, 13], is a software engineering methodology that encourages abstraction of the software complexities of service-based system architectures via models represented in the Unified Modeling Language (UML). The methodology encourages the separation of business logic from the platform and technical implementation concerns, allowing all to evolve independently [4, 5]. In 2003, Eric Evans coined the term Domain-Driven Design, a methodology that encourages system designers and architects to design software to reflect real-world business contexts and their users [6, 14]. While introduced before the microservice, DDD is a common design approach for microservice architectures today [8]. Even with the vast amount of documentation and research of these methodologies and with the well-defined guidelines for their adoption, challenges in identifying microservice boundary and granularity remain [2, 3, 7, 8]. Additionally, even though consequences of microservice design include an explicit interface positioned as the cornerstone of the design [15], challenges remain in defining precise published interfaces [1, 3], a critical aspect of design that is often considered late in the development cycle. An emerging methodology for microservice design, APIFirst Design, is gaining popularity as demonstrated by the available grey literature authored by industry practitioners. The fundamental principles of this emerging approach are that all business capabilities are exposed via an API and that the design of the system focuses on clear and welldefined APIs as the design foundation [7]. Stated benefits of API-First Design in addition to a well-defined API include increased exposure of business capabilities, improved identification of system components, clear boundaries between frontend and backend logic and the services themselves,
improved scale, and increased developer velocity due to parallelization of effort [8, 16, 17]. Next, this paper explores the state of API-First Design publications of the academic community, provides an overview of the grey literature available on the topic, and includes a discussion about future research opportunities that may advance the understanding and adoption of APIFirst Design.
1 Author
Nicole Beaulieu is currently the Vice President of Software Engineering for Figure’s lending engineering team and was previously employed as a Senior Software Engineering Manager at Intuit.
10.3
API-First Design
The importance of well-designed APIs is increasing as companies consider APIs as assets and commodities and therefore strive to maximize the capacities exposed via the suite of APIs [7,18]. Even with this increased attention, challenges regarding microservice boundaries and their APIs remain [1– 3, 7, 8]. Out of scope for this paper, it can be concluded that good API design is a critical aspect of modern software design and is a fundamental requirement of API-First Design. Publications targeting approaches for designing good APIs are plentiful; references include [18,19], and [20]. With APIFirst Design gaining attention in academia and emerging as an industry practice for system design, the following sections provide a review of the state of API-First Design in academia and a summary of API-First-focused grey literature published by industry practitioners.
10.3.1 State of Academia Prominent in practitioner-generated white papers and grey literature, API-First Design is a fledgling topic in academic research. Available publications generally include API-First design as a secondary theme. For example, in their 2019 paper [7], authors Wilde and Amundsen discuss the role that APIs play in the current digital landscape. They emphasize the important role of exposing capacities via an API in decentralized systems and the business for which they are designed. They further suggest that organizations with an API focus are striving to create an API Culture. In recognizing that API design is not an easy task, this paper focuses primarily on API design and management. The authors suggest that API-First Design is the preferred approach to exposing the capacities of a business via an API. The specifics of API-First Design are not discussed, leaving room to expand upon this topic in future research. Bennett et al. discuss in their 2021 article [21] a case study conducted with Saxo Bank. The study explored test suite definition and coverage of APIs exposed via an API-First approach to service design. API-First Design is a secondary topic of the paper; the article’s primary focus is test suite stability and functional coverage. The authors discover that
76
the challenges associated with testing are often a result of a lack of initial context (IC), a concept common in BehaviorDriven Development [22]. They conclude that the API-First approach leads to improved testing capabilities and practices due to increased collaboration in test plan definition as a result of the API-focused design methodology. In a noteworthy example of API-First research, authors Dudjak et al. present an API-First approach to designing a Backend as a Service (BaaS) platform [23]. Examples of a BaaS platform include data storage, user management, and file storage utilized by third-party application developers. The backend services, when combined by application developers, enable expedited delivery of full-featured platforms. In addition to the review of the design methodologies of BaaS components, the novelty of this research is found in the APIFirst approach adopted to design the system’s architecture and to establish the platform’s foundation. As discussed in the article, the benefits of this approach include loosely coupled services and faster time to market of products that utilize BaaS features. The authors of this paper recognize that the development process in an API-First approach begins with the definition and implementation of APIs that describe the platform; API-First Design frameworks are not included in the discussion. However, unlike the articles that reference API-First Design as a secondary topic, Dudjak et al. state the benefits of an API-First approach, including its facilitation of loose coupling between services and the representation of the services by a unique API. The goals and benefits included in this article coincide with those identified in industry and grey literature, a positive step toward closing the gap between the two sectors of the field. While not exhaustive, the articles discussed above represent literature published by the academic community at the time of writing. The minimal number of publications discussing API-First Design indicates the novelty of the research in academia and exposes an opportunity to explore the work of our colleagues in the industry.
10.3.2 Industry and Grey Literature In contrast to the limited research of API-First Design available in the academic setting, white papers, blogs, magazines, and video content from the industry discussing APIFirst design are plentiful. The extensive content provides an opportunity to understand the principles, benefits, and shortcomings of API-First Design. In 2002, Jeff Bezos, founder and CEO of Amazon, distributed a memo to his employees that mandated the definition and use of service interfaces for technology and interteam communication[24]. This API Mandate is an early rendition of an API-First approach to software design [25] and of building an API Culture. Fast forward nearly 20 years,
N. Beaulieu et al.
and Amazon offers a full suite of services and cloud-based products. Indicative of Amazon’s evolution as an API-First technology company since the 2002 Mandate are Amazon Marketplace Webservice (MWS) [26] and Amazon Web Services (AWS) [27], API suites that facilitate the use of Amazon’s marketplace and cloud-based technologies by entities external to the company itself. Another indicator of the importance of the API-as-a-commodity model is the API suite published for developer consumption by another Big 5 Tech Giant, Google [28]. Included in its API-accessed BaaS assets are Maps Platform, Google Ads, and Google Analytics [28]. A result of the digital transformation sparked by COVID19, authors Wang and McLarty [29] discuss the shift in mental model required for business leaders to support the technological and business reinvention necessary to endure such a transformation. The authors explain that the healthcare, banking, and logistics sectors may reap the benefits of API design and exposure. They also state that smaller companies may succeed amidst the more prominent companies by exposing their data via APIs. Similar to the Bezos API Mandate and the change in development and communication philosophy, the authors note that companies must learn to identify opportunities to expose business capacities via a suite of well-defined APIs [29]. In a 2020 blog post [17], author Chris Tozzi discusses the positive impacts of API-Driven development on an enterprise development strategy. Tozzi provides an invaluable summary of the benefits API-Driven [First] Design. Significant and relevant are the benefits of making the API the foundation of the architecture strategy. The approach ensures the exposure of components that may have been neglected in alternate API strategies also resulting in modular and cloud-friendly applications and services, improved CI/CD pipelines, and the assurance of compatible APIs due to the upfront focus on API development. Missing in academia, Tozzi summarizes steps that may assist in the successful adoption of API-First Design; a depiction of this process is shown in Fig. 10.2. The first steps represent an assessment of existing, new, and required APIs and the target and viable architectural patterns. The following steps focus on API development, testing, and monitoring integrated within the CI/CD pipeline. A final and essential step is the definition and adherence to a proactive API feedback loop. Based on the results of a 2018 public survey shown in Fig. 10.3, the Lazar article [30] indicates that of modern architectural approaches to distributed system design, architects are most in favor (67%) of API-First Architectures. In the article, Lazar identifies some benefits of API-First Design that may contribute to the surveyed architectural preference. The decoupling of a system’s frontend and backend components is included due to the well-defined API and clean dependencies. Furthermore, Lazar discusses API-
10
API-First Design: A Survey of the State of Academia and Industry
77
Fig. 10.2 API development cycle
First Design and its pluggable architecture. The modularity allows developers to focus on business logic and complexity rather than the application structure. Additionally, the naturally decoupled services enabled by the well-defined API increases opportunities for parallel development and an improved CI/CD pipeline [30]. Auth0 is a leader in user authentication service and provides a comprehensive API enabling integration with their services and platforms. In [11], based on experience at Auth0, Gontovnikas provides a high-level definition of API-First design, stating that the approach places the API at the center of software design, thus enabling services and software components to integrate easily. API-First Design discourages a code-first approach that leaves API attention to the end of a development cycle. Gontovnikas states that the benefits of the API-First Design approach include the ease with which software can be integrated and that the API is considered a valuable company asset. Both positively impact a company’s bottom line by improving time to market, customer experience, and increasing the adaptability of implemented software [11]. The anticipation of API-First Design extends beyond the high-tech industry and audience. For example, Huerta et al. discussed the positive impact that this design approach has on the news integration and knowledge distribution ecosystems [31]. The article highlights the benefits of API-First Design and the abstraction of implementation details behind a welldefined interface designed early in the development cycle. The authors suggest that this approach supports open innovation and open API concepts also backed by the tech industry [32, 33].
10.4
Future Research
The evolution of distributed and cloud-based technologies and the system architectures that support service-based products is leading a shift in business mindset that considers APIs an asset and commodity. In supporting this shift, software industry professionals are converging on the importance of quality APIs defined in an API-First Design approach [34]. While an imbalance in the number of publications exists between industry and academia, API-First Design is an emerging topic in the research setting. The infancy of API-First Design presents an opportunity to explore end-to-end the possibilities presented by this design methodology. A viable first step is a comprehensive survey of the grey literature available on the topic. Additional opportunities for study include the creation of a standard definition and the development of a common language and taxonomy that assist the software engineering community in discussing API-First Design. Opportunities also exist in leveraging the adoption strategies discussed in grey publications such as [17] and [30] to design a framework through which researchers and practitioners may adopt API-First Design. Research opportunities also exist in exploring the applicability of API-First Design. For example, model-based software generation is a well-researched topic that focuses on creating server-side code from a UML model. However, opportunities exist to automate UI/UX design and frontend code based on an API-First approach. This effort may be a natural extension to work performed by Rivero et al. in the
78
N. Beaulieu et al.
Fig. 10.3 Modern architectures survey, packt skill up 2018 [30]
MockAPI research [16]. A potential extension of Bennett et al. [21], additional research opportunities may exist in the examination of API-First Design and its potential benefit to the quality of fragile touch-points at the boundary of a microservice. The results may contribute relevant guidelines for test automation and quality and provide quantitative consideration of the impact on quality based on the early definition and testing of API contracts in a system. Finally, it is worth noting that tools and standards available in the software community may prove valuable in future research and exploration of an API-First Design approach. Among them are the OpenAPI Specification [33] and RESTful API Modeling Language (RAML) [35]. Using standards and tools that support API development and testing, such as Postman [36] and Swagger [32], may also increase the viability and applicability of this research to the industry.
10.5
is an emerging methodology for software architecture and design. Its fundamental tenet is that system design focuses on creating a clear and well-defined API as the foundation for the system design. The API foundation defines the integration and communication patterns of the combined microservices [23]. Stated benefits of API-First design include increased exposure of business capabilities, improved identification of services, a clear boundary between frontend and backend logic, improved scalability, and increased developer velocity due to the natural parallelization of efforts [8,16,17]. While API design and implementation are ubiquitous in software engineering, API-First Design as a methodology is in its infancy. Opportunities exist in creating a standard definition and vocabulary that describes the methodology, building a framework to guide its adoption, and exploring the qualitative and quantitative benefits of applying API-First Design to architectural system design.
Summary References
Microservice Architecture (MSA) is a standard design approach for distributed and cloud-based systems. However, attention to the interfaces through which services interact and are accessed is often an afterthought. As is evidenced by the quantity of grey literature about the topic, API-First Design
1. M. Fowler, J. Lewis, Microservices (2014). [Online]. Available: http://martinfowler.com/articles/microservices.html. Accessed 17 Oct 2021 2. J. Ghofrani, D. Lübke, Challenges of microservices architecture: A survey on the state of the practice, in ZEUS (2018)
10
API-First Design: A Survey of the State of Academia and Industry
79
3. J. Soldani, D.A. Tamburri, W.-J. van den Heuvel, The pains and gains of microservices: a systematic grey literature review. J. Syst. Softw. 146, 215–232 (2018) 4. O. MDA, Object management group model driven architecture (2008). Accessed 17 Oct 2021 5. I. Sommerville, Software Engineering, 10th edn. (Pearson, London, 2015) 6. E. Evans, Domain-Driven Design: Tackling Complexity in the Heart of Software (Addison-Wesley, Boston, 2004) 7. E. Wilde, M. Amundsen, The challenge of api management: API strategies for decentralized api landscapes, in Companion Proceedings of The 2019 World Wide Web Conference (2019). Accessed 17 Oct 2021 8. R.H. Steinegger, P. Giessler, B. Hippchen, S. Abeck, Overview of a domain-driven design approach to build microservice-based applications, in The Third International Conference on Advances and Trends in Software Engineering (2017) 9. C. Munns, Building API-driven microservices with amazon API gateway - AWS online tech talks (2018). [Online]. Available: https://www.youtube.com/watch?v=xkDcBssNd1g. Accessed 17 Oct 2021 10. H.R. Rothstein, S. Hopewell, Grey literature, in The Handbook of Research Synthesis and Meta-Analysis, vol. 2 (Russell Sage Foundation, New York, 2009), pp. 103–125 11. M.H. Valipour, B. Amirzafari, K.N. Maleki, N. Daneshpour, A brief survey of software architecture concepts and service oriented architecture, in 2009 2nd IEEE International Conference on Computer Science and Information Technology (2009), pp. 34–38 12. K.B. Laskey, K. Laskey, Service oriented architecture. WIREs Comput. Statist. 1(1), 101–105 (2009). [Online]. Available: https:// wires.onlinelibrary.wiley.com/doi/abs/10.1002/wics.8 13. I. Sacevski, J. Veseli, Introduction to model driven architecture (MDA), in Seminar Paper, University of Salzburg (2007) 14. V. Vernon, Domain-Driven Design Distilled (Addison-Wesley Professional, Boston, 2016) 15. L. Weir, Enterprise API Management: Design and Deliver Valuable Business APIs (Packt Publishing, Birmingham, 2019) 16. J.M. Rivero, S. Heil, J. Grigera, M. Gaedke, G. Rossi, MockAPI: An agile approach supporting API-first web application development, in ICWE, (2013) 17. C. Tozzi, The benefits of API-driven development and how you can implement it (2021). [Online]. Available: https://www.mertech. com/blog/api-driven-development. Accessed 17 Oct 2021 18. J.J. Bloch, How to design a good API and why it matters, in OOPSLA ’06 (2006) 19. J. Geewax, API Design Patterns. Manning (2021). [Online]. Available: https://books.google.com/books?id=XWU0EAAAQBAJ
20. J. Ofoeda, R. Boateng, J. Effah, Application programming interface (API) research: a review of the past to inform the future. Int. J. Enterp. Inf. Syst. 15, 76–95 (2019) 21. B.E. Bennett, A practical method for API testing in the context of continuous delivery and behavior driven development, in 2021 IEEE International Conference on Software Testing, Verification and Validation Workshops (ICSTW) (2021), pp. 44–47 22. M. Alférez, F. Pastore, M. Sabetzadeh, L.C. Briand, J.-R. Riccardi, Bridging the gap between requirements modeling and behaviordriven development, in 2019 ACM/IEEE 22nd International Conference on Model Driven Engineering Languages and Systems (MODELS) (2019), pp. 239–249 23. M. Dudjak, G. Martinovic, An API-first methodology for designing a microservice-based backend as a service platform. Inf. Technol. Control. 49, 206–223 (2020) 24. S. Yegge, Stevey’s google platforms rant (2011), https://gist.github. com/chitchcock/1281611, Accessed 17 Oct 2021 25. E. Wilde, Jeff bezos’ API mandate: What the five rules mean and do (2021). [Online]. Available: https://blog.axway.com/amplifyproducts/api-management/jeff-bezos-api-mandate. Accessed 17 Oct 2021 26. Develop for amazon selling partners (2021), https://developer. amazonservices.com/. Accessed 26 Nov 2021 27. Develop for amazon selling partners (2021). https://developer. amazonservices.com/. Accessed 26 Nov 2021 28. Google developers. https://developers.google.com/. Accessed 26 Nov 2021 29. T. Wang, M. McLarty, APIs aren’t just for tech companies (2021). [Online]. Available: https://hbr.org/2021/04/apis-arentjust-for-tech-companies. Accessed 17 Oct 2021 30. A. Lazar, 8 reasons why architects love API driven architecture (2018). https://hub.packtpub.com/architects-love-api-drivenarchitecture/. Accessed 17 Oct 2021 31. J. Huerta, C. Childs, Accelerating news integration in automatic knowledge extraction ecosystems: an API-first outlook (2015). ArXiv abs/1509.02783 32. Swagger.io, API development for everyone (2021). [Online]. Available: https://swagger.io/. Accessed 17 Oct 2021 33. Openapis.org, Openapi initiative (2021). [Online]. Available: https://www.openapis.org/. Accessed 17 Oct 2021 34. L. Murphy, M.B. Kery, O. Alliyu, A.P. Macvean, B.A. Myers, API designers in the field: Design practices and challenges for creating usable APIs, in 2018 IEEE Symposium on Visual Languages and Human-Centric Computing (VL/HCC) (2018), pp. 249–258 35. Raml.org, The simplest way to model APIs (2021). [Online]. Available: https://raml.org/. Accessed 17 Oct 2021 36. Postman.com, Postman API platform (2021). [Online]. Available: https://www.postman.com/. Accessed 17 Oct 2021
Part II Data Science & Engineering
A Quality Dimension Analysis of Open Government Data in Undergraduate Public Funding in Brazil
11
Marcelo Moreira West and Glauco de Figueiredo Carneiro
Abstract
This paper describes the conduction of a tutorial to characterize the experience with dealing with nonconformities of open government data. The goal was to analyze data quality to identify limitations and errors relative to the use of open government data from the viewpoint of researchers and users in the context of Brazilian Student Financing (FIES). We planned the tutorial in seven steps to analyze the FIES data quality dimensions and present the results of difficulties and challenges faced by the participants. Keywords
Open government data · Data quality · Data science
11.1
Introduction
Open government data has attracted ever-increasing attention from citizens due to its relevance to understanding several aspects of life in towns and cities. Governments’ bodies worldwide have provided data that is of the main interest of citizens. The literature argues that adopting best practices to provide open data to citizens has the following goals: promoting transparency, encouraging participatory governance, and identifying government accountability for decisions. Providing open government data without considering relevant quality dimensions is a significant constraint to enhancing its practical use. Incomplete, imprecise, inconsistent, and M. M. West Instituto Federal Baiano (IFBAIANO), Salvador, Bahia, Brazil e-mail: [email protected] G. d. F. Carneiro () Universidade Salvador (UNIFACS), Salvador, Bahia, Brazil e-mail: [email protected]
duplicate data are just examples of nonconformities in this scenario. Studies have described the increased use of open government data portals in different regions of the world [1]. For example, there is a representative increase in datasets published on the Australian government portal in the order of 900% between 2013 and 2015 [1]. In the same period, the New Zealand government portal reached the milestone of 3,800 datasets, United Kingdom 23,000, United States 194,000, and Canada 240,000 [1]. Open government data can contribute to the development of aware and active citizens [2]. However, the lack of data quality is a challenge to achieving this goal. For example, there is a challenge to deal with the non-standardization of data formatting, and the non-conformity in the provision of metadata [3]. The tutorial presented in this article aims to analyze data quality to identify limitations and errors regarding the use of open government data from researchers and users not originally from information technology in the context of public funding to undergraduate students. This paper is structured as follows: Sect. 11.2 presents the main concepts of data quality dimensions. Section 11.3 details the methodological procedures for planning and conducting the tutorial. Section 11.4 presents further discussion regarding the findings in the context of open government data. Finally, Sect. 11.5 discusses the challenges faced in analyzing open government data.
11.2
Data Quality Dimensions
Analyzing data quality dimensions makes it possible to classify and identify data to achieve users’ goals [4]. These dimensions contribute to selecting a specific dataset by potential users, which directly impacts the analysis results. The challenges to dealing with open government data influence
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_11
83
84
M. M. West and G. d. F. Carneiro
Table 11.1 Study goal according to the GQM approach Analyze for the purpose of with respect to the from the point of view of in the context of
data quality identifying limitations and errors use of open government data researchers and users Brazilian Student Financing (FIES)
the decision-making process due to data non-conformities. We present a set of quality dimensions already discussed in the literature and their description. Interpretability assesses whether it is possible to interpret the meaning and properties of data sources through the provided documentation and metadata [5]. This dimension highlights the importance of data properties to the practical use of data meaning. Comprehensibility evaluates how easily the data is understood [6]. In this case, the better the comprehensibility, the less effort we need to understand data. Credibility evaluates how genuine and trustworthy the data is [6]. Otherwise, it does not correspond to reality. Integrity assesses whether data is missing and whether it has sufficient breadth and depth to use [7]. We can observe a relationship between credibility and integrity, where integrity directly affects the data credibility. Accuracy evaluates whether the data is reliable and error-free [5]. The Concise Representation evaluates to what extent the data is represented in a compact form [7]. The Consistency evaluates whether there is compatibility between different data values [8] and completion evaluates the number of filled attributes [9]. Much of the open government data has errors such as missing attribute values, incorrect attribute values or different representations in the same dataset [10]. The missing attribute values correspond to the non-completion of mandatory attributes [10]. The incorrect values are considered those that do not correspond to the real situation [10]. The different representations occur when the attribute values have different formats [10]. The illegal values are considered those outside the domain of valid values [10]. To reach correct conclusions, the data must have a high level of quality [10]. In the following, we describe common errors usually found in data. The first is a set of outliers that deviate from the distribution of values. Duplicate data referring to the same entity also occur as a common error. Other occurrences are values that violate integrity constraints such as nullity and uniqueness [11].
11.3
Methodology
This section describes the tools we adopted to conduct the tutorial. Then we introduce the execution steps. Table 11.1 presents the goal of the study according to the Goal Question Metric (GQM) approach [12]. To perform the activities, the participants used Google Meet, a service that allows individual or collective interaction
Table 11.2 Tutorial steps performed by the participants Step Details 1 Getting a copy of the office tools package installation file Libreoffice and running 2 Obtaining the compressed file with the FIES data in ”RAR format available on the platform Google Classroom 3 Unzip the file obtained in the previous step 4 Validation of the unzipped file, comparing its size with with the file size available on the Brazilian Open Data Portal 5 Importing FIES data into the Libreoffice Calc tool 6 Analysis of FIES data quality dimensions 7 Sending the questionnaire APPENDIX A answered with the aid of the Data Dictionary ANNEX I to the organizers
through [13] videoconferences. They also used the Google Classroom, a component of the Google for Education application package made up of the (gmail) mail manager, (drive) file repository and text editor, spreadsheets and presentations (docs) [14]. The participants analyzed data provided by the Student Financing Fund (FIES), available on the Brazilian Open Data Portal in ”CSV” format.1 The reason to use data from the second half of 2019 was that they were the most recent available data in the portal. (A) Planning The FIES data quality dimensions analysis followed the seven steps presented in Fig. 11.1. The first step consisted in the installation of the Libreoffice tool. The second step orresponded to the download of the FIES open data file in a compressed format available on the platform Google Classroom. In the third step, participants uncompressed the file, and in the fourth step, they validated the content of the unzipped file, comparing its size with the file size available on the Brazilian Open Data Portal. In the fifth step, participants opened the unzipped file into the Calc tool in Libreoffice. In the sixth step the participants performed the data analysis, and in the seventh step they filled in a questionnaire. Table 11.2 presents the steps of the tutorial. Selection of Participants We selected participants based on the following criteria: not having training in the Information Technology area, having skills in the use of computers, software, internet access, having an e-mail account on the google platform, and using the operating system (Windows or Linux). Dimension Selection The criteria for selecting the data quality dimensions were: dimensions related to data usage and decision making.We choose the following dimensions: concise representation, interpretability, understandability, credibility, integrity, and accuracy. 1 http://portal.mec.gov.br
11 A Quality Dimension Analysis of Open Government Data in Undergraduate Public Funding in Brazil
85
Fig. 11.1 Tutorial steps
Results and Discussion of Results
aimed at identifying errors as reported in Table 11.3. Table 11.6 shows the number of errors found in the attributes of the FIES data by analyzing the responses of the participants. The evidence found were attributes filled in with spelling errors, filled in with illegal values, or not filled in. In The third step, the participants registered the errors they detected. In contrast, the fourth step aimed at proposing suggestions for improving the quality of data represented in Table 11.5. Data analysis and error identification considered the following dimensions of data quality: interpretability, understandability, concise representation, credibility, integrity and accuracy. It is essential to highlight that the data published in open government portals must meet the principles of the selected dimensions, as they address that the users should easily interpret data. Moreover, complete, error-free, and reliable data contributes to an effective decision-making process (Table 11.7). The organizers analyzed the answers provided by the participants participants during the identification of errors in the data. Participants answered the questionnaire provided in Table 11.6 supported by information provided by Table 11.8. Table 11.3 presents the list of errors found by the participants.The following errors were identified:
This section discusses the results of the FIES data quality analysis and also identifies opportunities for further improvements (Tables 11.3, 11.4, and 11.5). The analysis of the quality of the FIES data revealed nonconformities as presented in Table 11.6. Figure 11.2 presents the four stages of the analysis. The first step consisted of reviewing the literature to identify frequent errors represented in Table 11.3. The second step
Missing Attribute Value They were found in the attributes Abbreviation of the Federation Unit, Name of the Municipality, Number of the National Register of the Legal Entity of the Maintainer, Name of the Corporate Name of the Maintainer, Code of the Municipality of the Maintainer and Name of the Municipality given in blank, which demonstrates noncompliance with the dimension completeness, according to questions 1 and 18.
(B) Execution Figure 11.1 presents the steps performed by the participants. The participants took part in the tutorial in two sections. The organizers presented the proposed activities to download, install, and configure the environment in the first section. In the sequence, the participants performed the steps. The participants could ask the organizers questions regarding the activities. Participants performed seven steps to analyze the data quality dimensions: In the first, they obtained a copy and installed the Libreoffice tool. In the second, they obtained a copy of the FIES open data in a compressed file. In the third, they unzipped the file copied in the previous step. On the fourth, they validated the extracted file. The following tools were used by the participants: operating system Windows 10, browser Google Chrome Version 95.0.4638.54 64-bit and spreadsheet editor LibreOffice Community Version 7.2.0.4 (x86), WhatsApp Version 4.1., file unzipper Winrar Version 6.0 64 bits and mail manager.
11.4
86
M. M. West and G. d. F. Carneiro
Fig. 11.2 Tutorial data analysis steps
Table 11.3 Frequent quality mistakes in open data Error identifier E1
Error type Incorrect values
E2
Incorrect value
E3
Incorrect values
E4
Incorrect values
E5
Incorrect values
E6
Incorrect values
E7 E8
Missing attribute values Incorrect values
E9
Illegal values
Error name Data used for a purpose other than intended Text type data filled with numeric type Numeric type data filled with text type Date-type data filled in with a date greater than or equal to 100 years from the current date Data filled with spelling errors Data filled with more than one attribute Unfilled data Data filled with undecipherable characters Data filled with illegal values
Error description E-mail attribute filled in with a name Attribute name filled with cpf number Phone attribute filled in with name Attribute date of birth filled with date of next century
County attribute filled in without county=”Sao Paulo” accent but should be accented Attribute name filled in with name=”Maria da name and process number Silva Souza 99393346-20” Blank corporate name attribute corporate name=”” County attribute filled with special symbols from the ASCII table Attribute monthly value filled with negative value
Table 11.4 Distribution of error quantity by error type Error type Incorrect values Missing attribute value Illegal values Total
Number of 385949 2697 36 388682
Error example e-mail=”Maria da Silva Souza name=”999.999.99999” phone=”Maria da Silva Souza” birth date=2121-01-01
Error frequency – n(%) 99,297 0,694 0,009 100
Incorrect Values In the attributes Type of Contracted Surety, Sponsor’s Municipality Code, Higher Education Institution Municipality Code, and Student Residence Municipality Code, data was represented in undecipherable characters and the Maintainer’s Corporate Name, Education Institution Name attributes Higher (NO IES), Name of Higher Education Institution (NO IES EXT ALUNOS)
Dimension of associated quality Consistency Consistency Consistency Accuracy
Accuracy concise representation Completeness
county=”S – o Paulo” Interpretability, Understandability e Credibility monthly =-820,41 Consistency e Credibility
Table 11.5 Improvement suggestions No. 1 2 3
Suggestion Make attributes found in blank mandatory Use the ”Western Europe (Windows 1252/Latin1) character setting” Include missing attribute descriptions in the data dictionary
data with an incorrect accent, which demonstrates noncompliance with the dimensions: interpretability, understandability, credibility and accuracy, according to the questions 14 and 20. Illegal Values Invalid values were found in the Monthly Value attribute, which demonstrates non-compliance with the data quality dimensions consistency and credibility.
11 A Quality Dimension Analysis of Open Government Data in Undergraduate Public Funding in Brazil
87
Table 11.6 Distribution of error quantity by attribute Attributes SG UF NO MUNICIPIO CO MANTENEDORA NU CNPJ MANTENEDORA NO RAZAO SOCIAL MANTENEDORA CO MUNICIPIO MANTENEDORA CO IES NO IES CO MUNICIPIO IES CO PROCESSO CO CONTRATO FIES CO ADITAMENTO CO AGENTE FINANCEIRO NO AGENTE FINANCEIRO SG AGENTE FINANCEIRO NU MES NU SEMESTRE NU ANO VL MENSALIDADE CO INSCRICAO TP FIANCA NU ANO EXERCICIO INSC NU PERCENTUAL PROUNI NU PERCENT SOLICITADO FINANC VL REPASSE DT NASCIMENTO ST DEFICIENCIA ST ENSINO MEDIO ESCOLA PUBLICA CO CIDADE SG SEXO DS SEXO CO ESTADO CIVIL DS ESTADO CIVIL SG RACA COR DS RACA COR CO INSCRICAO EXT ALUNOS CO CONTRATO FIES EXT ALUNOS CO IES EXT ALUNOS NO IES EXT ALUNOS CO CAMPUS NO CAMPUS VL PERC FINANCIAMENTO VL MENSALIDADE EXT ALUNOS QT SEMESTRE FINANCIADO CO ADITAMENTO EXT ALUNOS CO CURSO DS CURSO CO TIPO CURSO DS TIPO CURSO CO PERIODICIDADE CUR DS PERIODICIDADE CUR VL SEMESTRE NU ANO PROC NU MES PROC Total
E1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
E2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
E3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
E4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
E5 0 0 0 0 375122 0 0 975 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 975 0 8877 0 0 0 0 0 0 0 0 0 0 0 0 0 385949
E6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
E7 351 351 0 351 351 351 6 6 366 0 0 0 0 0 0 0 0 0 0 0 0 6 126 6 0 0 0 0 0 0 0 123 123 0 0 0 0 0 0 0 0 6 0 6 0 0 0 33 33 51 51 0 0 0 2697
E8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
E9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 36 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 36
Subtotal 351 351 0 351 375473 351 6 981 366 0 0 0 0 0 0 0 0 0 36 0 0 6 126 6 0 0 0 0 0 0 0 123 123 0 0 0 0 0 975 0 8877 6 0 6 0 0 0 33 33 51 51 0 0 0 388682
88
M. M. West and G. d. F. Carneiro
Table 11.7 Questionnaire 01
02
03 04
05 06 07
08
09
10 11
12 13 14
15
16
17
18 19
20
In your opinion, the registration data which contains the sponsor code equal to 17024, the FIES contract code equal to 2187276, the reference month of the payment statement equal to 8 and the reference year of the payment statement equal to 2019 are correct? You assume that the record data contains the maintainer code equal to 1534, the FIES contract code equal to 274338, the reference month of the payment statement equal to 5 and the reference year of the payment statement equal does 2012 have inconsistencies? Do you believe that it is possible to identify the names of the municipalities in the contract data number 2830492? Do you think there is no flaw in the record that contains the municipality of Canoas, the sponsor code equal to 314, the contract code 135498, the reference month of the statement equal to 1 that prevents the identification of the financial agent? Do you consider that in the municipality of Alagoinhas, sponsor code 2079, there are 3 students with disabilities? Do you consider that in the contract code equal to 2165590, registration number equal to 3966632, is it possible to identify that the candidate is indigenous? Is it correct to state that in the contract code equal to 2378896, enrollment code equal to 2945086, whose name of the higher education institution is the same as Facul- dade da Amazo^nia Ocidental, is the course description architecture? In your opinion, the data of the institution code equal to 780, contract code equal to 2286193, maintaining code equal to 14514, high school status in public school equal to S allow to identify whether the student studied high school in the public network ? Do you consider that in the contract code equal to 2128687, enrollment code equal to 3068960, reference month of the payment statement, the data of the city code of the educational institution and the city code are redundant? Do you agree that in the municipality equal to Arapiraca, description of brown race we have three candidates registered? Have you found field(s) used with a different purpose than expected? Example of problem found: e-mail field filled in with name. If so, enter the names of the fields found that were used for a different purpose than expected: Did you find alphabetic field(s) filled with numbers?Example of problem found: field name filled with cpf. If so, enter the names of the alphabetic fields that have been filled in with numbers: Did you find numeric field(s) that were filled with text?Example of problem found: phone field filled in with number and name. If so, enter the names of the numeric fields that have been filled in with text: Did you find field(s) filled in with spelling errors? Example of problem found: field no razao social mantenedora is incorrectly written “educacao”. If so, enter the names of the fields found filled with spelling errors: Did you find field(s) filled in with a date that is more than 100 years apart from the current date? Examples of prob- lems found: field dt nascimento filled with 01/01/1900 or 01/01/2100. If so, enter the names of the fields found filled in with a date greater than 100 years from the current date: Did you find records that represent the same data, but with different identifications? Example of problem en- countered: name filled in one record in complete form and in another record in abbreviated form. If so, inform the number of records found that represent the same data, but with different identifications: Did you find field(s) filled with more than one attribute? Example of problem encountered: name field filled with process name and number If so, enter the names of the fields found filled with more than one attribute: Did you find unfilled field(s)? Example of problem found: field no razao social maintainer not filled. If so, enter the names of the fields found without filling it in: Did you find field(s) with name(s) different from the names in the data dictionary? If so, answer: What fields did you find with names different from the names in the data dictionary? What do you recommend to be done to correct the names of fields that were found with names different from the names in the data dictionary? What did you do to correct the field names that were found with different names from the data dictionary names? Did you find field(s) with data with undecipherable characters? If yes, please answer: What fields did you find with data with undecipherable characters? What do you recommend to be done to fix the fields you found with data with undecipherable characters? What did you do to fix the fields found with data with undecipherable characters?
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
( ) Yes
( ) No
11 A Quality Dimension Analysis of Open Government Data in Undergraduate Public Funding in Brazil Table 11.8 Data dictionary ANNEX I – DATA DICTIONARY Field SG UF NO MUNICIPIO NU CNPJ MANTENEDORA NO RAZAO SOCIAL MANTENEDORA CO MUNICIPIO MANTENEDORA CO IES NO IES CO MUNICIPIO IES CO PROCESSO CO CONTRATO FIES CO ADITAMENTO CO AGENTE FINANCEIRO NO AGENTE FINANCEIRO NU MES NU SEMESTRE NU ANO VL MENSALIDADE CO INCRICAO TP FIANCA NU ANO EXERCICIO INSC NU PERCENTUAL PROUNI NU PERCENT SOLICITADO FINANC VL REPASSE DT NASCIMENTO ST DEFICIENCIA ST ENSINO MEDIO ESCOLA PUBLICA CO CIDADE SG SEXO DS SEXO CO ESTADO CIVIL DS ESTADO CIVIL SG RACA COR DS RACA COR CO INSCRICAO EXT ALUNOS CO CONTRATO FIES EXT ALUNOS CO IES EXT ALUNOS NO IES EXT ALUNOS CO CAMPUS NO CAMPUS VL PERC FINANCIAMENTO VL MENSALIDADE EXT ALUNOS QT SEMESTRE FINANCIADO CO ADITAMENTO EXT ALUNOS CO CURSO DS CURSO VL SEMESTRE
Description Federation Unit Acronym City Name National Register of Legal Entity (CNPJ) of the sponsor Name of the sponsor’s corporate name Municipality code of the sponsor according to IBGE Code of the Higher Education Institution (IES) in the e-MEC electronic system Name of Higher Education Institution – IES Municipality Code of Higher Education Institu- tion IES according to IBGE Reference code for the payment process in the FIES Computerized System (SisFIES) Contract code in the FIES Computerized System (SisFIES) Identification code of the amendment to the Sis-Fies financing contract. Code of the Financial Agent responsible for student financing: (1) Banco do Brasil (2) Caixa Economica Federal. Name of the Financial Agent. Payment statement reference month. Reference semester of the payment statement. Payment statement reference year. Monthly fee for the financed course. Student registration code in the FIES Computer- ized System (SisFIES). Type of surety contracted: Normal / Conventional (N); Educational Credit Operations Guarantee Fund – FGEDUC (F) and Solidarity (S). Year of exercise referring to the student’s en- rollment in the FIES Computerized System (Sis- FIES). Scholarship percentage related to the University for All Program – ProUni. Percentage of funding requested by the student. Total amount of financing transferred to the spon-sor. Student’s date of birth. The student has special needs: No (N) and Yes (Y). The student attended high school in a public school: No (N), Yes (S) and Partially (P). Code of the student’s residence municipality ac- cording to IBGE. Gender Acronym: (F) Female (M) Male. Description of sex. Marital Status Code: (1) Single (2) Married (3) Widowed (4) Separated (5) Divorced (6) Stable Union. Description of marital status. Acronym Race/Color. Description Race/Color (A) Yellow (B) White (I) Indian (N) Black (P) Brown. Student registration code in the FIES Computer- ized System (SisFIES). Contract code signed with the Student Financing Fund – FIES. Code of the Higher Education Institution (IES) in the e-MEC electronic system. Name of Higher Education Institution – IES. Campus code in the e-MEC electronic system. Campus name. Percentage of funding granted to the student. Monthly fee for the financed course. Number of semester financed. Identification code of the amendment to the fi-nancing contract in the FIES Computerized Sys- tem (SisFIES). Course code in the e-MEC electronic system. Course description. Semester fee of the financed course
89
90
M. M. West and G. d. F. Carneiro
The following attributes were found without a description in the data dictionary: Course Type Code, Course Type Description, Course Frequency Code, Course Frequency Description, Process Year, Process Month and Financial Agent Acronym. The participants found it challengingt to understand and interpret the attribute that refers to the year of the payment statement and were unable to identify whether the student is indigenous through the attribute Description Race/Color, as per questions 2 and 6. Based on the results obtained in this study, we summarize improvement suggestions in Table 11.5. The first focuses on avoiding blank data records. The second argues the benefits of preventing indecipherable data, and the third indicates the advantages of using data to facilitate the content’s understanding and interpretation.
11.5
Conclusion
At the beginning of the article, there was doubt about the limitations faced by users of data published in open portals, so it became essential to study the analysis of the dimensions of data quality to know these limitations and know how to overcome them. Therefore, the article aimed to analyze the quality of data to identify limitations and errors about using open government data from the perspective of researchers and users in the context of Brazilian Student Financing (FIES). It appears that the general objective was met, as effectively, with the completion of the tutorial, it was possible to find quality problems in the data and propose solutions. The first goal to be met was the preparation of the environment to carry out the data quality analysis, in which the participants successfully achieved this goal, as they performed the installation of support software, obtained a copy of the FIES data, and imported them for the tool used in the analysis. The second goal to be met was the analysis of the dimensions of data quality. The participants successfully reached this goal, as we conducted the study in total and the evidence tabulated. The third goal to be met was elaborating a list of suggestions for improvements. We selected users who did not have training in the Information Technology area. The quality dimensions we focused on were: interpretability, understandability, concise representation, credibility, integrity, and accuracy-related to the use of data.
Given the proposed methodology, it is clear that the tutorial could have been carried out with other user profiles, since in this work, given the limited time, it was only possible to do it with an audience without training in exact sciences. Thus, it is recommended for future work to develop a tutorial focusing on users trained in data science so that this profile of users can consider data migration to relational and nonrelational databases.
References 1. S. Sadiq, M. Indulska, Open data: Quality over quantity. Int. J. Inf. Manage. 37(3), 150–154 (2017) 2. A.d.A.P. Silva, D.A.A. Monteiro, A. de Oliveira Reis, Qualidade da informac¸a˜o dos dados governamentais abertos: Análise do portal de dados abertos brasileiro. Revista Gesta˜o em Análise 9(1), 31– 47 (2020) 3. L. Ding, V. Peristeras, M. Hausenblas, Linked open government data [guest editors’ introduction]. IEEE Intell. Syst. 27(3), 11–15 (2012) 4. F.G. Dutra, R.R. Barbosa, Modelos e critérios para avaliac¸a˜o da qualidade de fontes de informac¸a˜o: uma revisa˜o sistemática de literatura. Informac¸a˜o & Sociedade. 27(2) (2017) 5. C. Batini, B. Pernici, Data quality management and evolution of information systems, in 19th IFIP World Computer Congress – WCC 2006, vol. 214 (2006), pp. 51–62 6. L.L. Pipino, Y.W. Lee, R.Y. Wang, Data quality assessment. Commun. ACM 45(4), 211–218 (2002) 7. R. Vaziri, M. Mohsenzadeh, A questionnaire-based data quality methodology. Int. J. Database Manage. Syst. 4(2), 55 (2012) 8. M. Scannapieco, T. Catarci, Data quality under a computer science perspective. Archiv. Comput. 2, 1–15 (2002) 9. K.J. Reiche, E. Hofig, I. Schieferdecker, Assessment and visualization of metadata quality for open government data, in CeDEM 2014, International Conference for E-Democracy and Open Government (2014), pp. 335–346 10. P. Oliveira, F. Rodrigues, P. Henriques, Limpeza de dados-uma visa˜o geral. Data Gadgets (2004), pp. 39–51 11. Z. Abedjan, X. Chu, D. Deng, R.C. Fernandez, I.F. Ilyas, M. Ouzzani, P. Papotti, M. Stonebraker, N. Tang, Detecting data errors: Where are we and what needs to be done? Proc. VLDB Endow. 9(12), 993–1004 (2016) 12. V.R. Basili, H.D. Rombach, Towards a comprehensive framework for reuse: A reuse-enabling software evolution environment (1988) 13. D.d.F.F.A. Sant, D.V. Sant’Anna, et al., Google meet como modalidade de ensino remoto: Possibilidade de prática pedagógica. Anais do CIET: EnPED: 2020-(Congresso Internacional de Educac¸a˜o e Tecnologias – Encontro de Pesquisadores em Educac¸a˜o a Dista^ncia) (2020) 14. R.H.N. Diniz, J.C.F. de Almeida, B.F.L. Rodrigues, M.M.R. Marmol, E. Superior, Utilizando o google classroom como ferramenta educacional–percepc¸o˜es e potenciais.
A Survey of Real-Time ETL Process Applied to Data Warehousing Environments
12
Flávio de Assis Vilela and Ricardo Rodrigues Ciferri
Abstract
ETL (Extract, Transform, and Load) is an essential process required to perform data extraction from data sources, transforming the extracted data into an integrated format and loading the integrated data into a data warehouse. Moreover, the ETL process can be performed in a realtime way. This paper presents a survey of real-time ETL process applied to data warehousing environments. The related studies were collected by means of a systematic review conducted by a methodological protocol. Moreover, we grouped the studies by categories, which had not been made before in the literature. The results showed that we have a wide range of opening opportunities of original research, as there are a lot of initial ideas and in turn they have not yet been tested, validated, and compared to related work. Keywords
Survey · ETL · Real-time · Data warehouse · Data warehousing
12.1
Introduction
Nowadays, the collecting of data that support the decisionmaking process has become essential for different applications domains, such as health care systems, road control F. d. A. Vilela () Department of Computing Federal Institute of Goiás, Jataí, GO, Brazil e-mail: [email protected] R. R. Ciferri Department of Computing (DC), Federal University of São Carlos (UFSCar), São Carlos, SP, Brazil e-mail: [email protected]
systems, and digital agriculture. These data are located in operational sources (OLTP systems). Furthermore, the operational data are available in different formats, such as relational databases, semi-structured and unstructured sources, in text files, XML or JSON files, and spreadsheets [1,2]. The decision-making process is performed based on the data of interest gathered from the operation sources. Therefore, to provide value for the operational data and effectively aid the decision-making process, the ETL (Extract, Transform, and Load) process should be applied. This process is responsible for dealing with the operational data and performing the extraction, cleaning, processing and integration of data. Thus, the data of interest available in the operational environment are sent to the data warehousing environment and stored into a dimensional, homogeneous and integrated database named as data warehouse (DW) [1, 3, 4]. Traditional ETL process is composed and represented by a workflow of tasks. These tasks are logically split in three major steps: (1) The first one is called Extraction (E). The main goal is to extract or capture the data of interest to the decision-making process from its respective operational sources. (2) The next step is called Transformation (T), where the data cleaning, data processing and data integration is performed. (3) The third step is called Loading (L), which is responsible for loading data to the DW [5,6]. The result of the ETL process is to make the data of interest of the operational environment available to the decision-making process, that is, storing operational data into the DW in an integrated and consistent way [5, 7]. In the traditional ETL process, the process for gathering data from the operational sources to the DW frequently takes place in a specific and predefined period of time, on a loading time window or according to the organization’s business rules [1, 5, 8]. In fact, the data can be gathered once a day, once a week or even once a month. In this time, the operational and the information environments are inconsistent. Besides, when using the loading time window for gathering data, both
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_12
91
92
operational and informational environments should be offline to perform the process. To perform the ETL tasks, some Change Data Capture (CDC) techniques are applied into the ETL process to aid the extraction phase. These techniques can be a trigger, log file, delta file, snapshot or replication. In this way, the ETL process is highly dependent on operational sources. On the other hand, a lot of applications have emerged and these applications need to gather operational data from operational sources in a real-time way, such as health care systems [9] and systems that handle data from sensors [10]. For this purpose, the ETL process has been performed in a higher frequency than traditional ETL process or in each idle time of the application so that it can simulate the near realtime approach [3, 5, 11]. In some cases, the operational data should be available in real-time into the DW to the decisionmaking process [12,13]. As a consequence, the requirements for a real-time environment, that is, availability, scalability and low response time, must be ensured by the ETL process [12]. In recent years, some literature reviews were proposed in the literature which focuses on: (1) available techniques to be applied to real-time ETL process [8, 14], (2) all features of real-time ETL process into a data warehousing environment [4], and (3) some architectures to make a data warehousing with real-time ETL capabilities [15]. However, there is no work that classifies all general studies and their applied purposes. Moreover, they did not discuss how the related work deals with the requirements of a real-time environment, nor which are the techniques applied to make possible the real-time ETL process and which CDC techniques are used in the ETL process. By solving this gap, this paper provides a literature review of the studies that deal with real-time ETL process applied to data warehousing environments. The related work is separated into two groups: the first one is related to investigations whose main goal is a general study; the second one is regarding studies that focus on applied purposes, that is, architecture, methodology, methods and applications. By the end of this paper, we show a comparison of all work and we point out which are the state of the art for the real-time ETL process. This paper is organized as follows: Sect. 12.2 summarizes the most important studies that was proposed in the literature, Sect. 12.3 shows and discusses the ETL process and its phases, Sect. 12.3.3 discusses the methodology used to obtain the related work, Sect. 12.5 depicts the comparison among the related work and Sect. 12.6 concludes the article.
F. d. A. Vilela and R. R. Ciferri
12.2
Related Work
Many investigations have been accomplished in the literature to enhance the freshness of data into the DW by applying the ETL process. In the same way, some literature reviews were proposed to highlight the main architectures to perform realtime ETL process in data warehousing environments. Vassiliadis and Simitsis [16] proposed a literature review of traditional ETL process as well as its issues and challenges and highlighted all aspects to perform the ETL process in a near real-time ETL process. Kabiri and Chiadmi [17] showed some commercial and open source tools and some prototypes to perform ETL process, beyond to show the design of the conceptual model of the ETL process. Vassiliadis [18] proposed a literature review from the conceptual and logical point of view of the ETL process. From the point of view of general study to perform ETL process in the real-time way, Sabry and Ali [4] showed an overview of a basic DW architecture as well as an overview of the ETL process and the main techniques to apply it into realtime data warehousing environments. However, the authors did not show architectures to perform the ETL process for real-time data warehousing environments. Bouaziz et al. [15] proposed a literature review which showed the main techniques and two other architectures to perform ETL process in a real-time way. However, the authors just showed the whole architecture and they did not group them by each phase of the ETL process. Sabtu et al. [19] showed an overview of the ETL process grouping the study on extraction, transformation and loading phases, and also they showed the real-time requirements that should be respected into a real-time ETL process. However, the authors just showed a literature review and they did not show an applied purpose to be used into a real-time ETL process. As we can note, a lot of literature reviews have been conducted to show all related work regarding the real-time ETL process in data warehousing environments. However, the previews comparative studies did not show the following features: (1) a comparison of the related work with the characteristics of a real-time environment; (2) an indicative which was the approach of the each paper, that is, if it was an architecture, method, methodology, strategy or application; (3) they did not show which CDC technique that was used; (4) they did not show which techniques are applied to make the ETL process with real-time capabilities; (5) they did not show the state-of-the-art of each phase of ETL process. So, in this paper we present a comparison of all general studies
12
A Survey of Real-Time ETL Process Applied to Data Warehousing Environments
and applied investigations of the real-time ETL process into data warehousing environments. Finally, we will show a comparison of all related work and point out the state-of-theart for real-time ETL process applied to data warehousing environments.
12.3
ETL Process
The ETL process is an inseparable part of the data warehousing environment. Kabiri and Chiadmi [17] says that about 70% of the whole cost of developing a data warehousing environment is spent to design and build the ETL process. This process is composed and represented by a workflow of tasks. These tasks are logically split in three major steps:
12.3.1 Extraction (E) The first step is called Extract, which the main goal is to extract the data of interest to the decision-making process from its respective operational sources. In this step, it connects to the operational sources from the operational environment in order to gather the data of interest. To do this, this step needs to know how to connect to the operational sources in order to access the operational data, as well as it needs to deal with the heterogeneity of the whole sources of the operational environment.
12.3.2 Transformation (T) The next step is called Transformation, where the data cleaning, data processing and data integration is performed. From this step, the whole ETL process makes value to the extracted data. In other words, this step focuses on making available accurate data, that is, the data should be correct, unambiguous, consistent and complete.
12.3.3 Loading (L) The last step is called Loading, which is responsible for loading data to the DW [5, 6]. The result of the ETL process is to make the data of interest of the operational environment available to the decision-making process, that is, storing operational data into the DW in an integrated and consistent way [5,7]. In this step, the data can be stored into DW in two ways: (1) bulk-load, that is, all data from operational sources are stored into DW every time the ETL process is performed; (2) incremental loading, that is, the data from operational sources are stored into the target DW periodically, in any time interval or continuously [6].
12.4
93
Methodology
To make it possible to find all related work for the real-time ETL process, we made a systematic review, which the main goal was to identify the architectures, strategies, techniques, methodologies, applications and general studies to allow us getting an overview of all characteristics of the real-time ETL process. The systematic review was guided by a protocol, which means that we made use of a method to allow us to collect a wide range of related work that deal with real-time ETL process. The protocol was composed of the objectives, the research questions, the sources of search, the keywords, the selection criteria of inclusion and exclusion of a work, the idiom of the studies and finally the search strings. We used Parsifal software (https://parsif.al/) to support the systematic review. For the sources of search, we searched IEEEXplore Digital Library, ACM Digital Library, Scopus, Web of Science and Google Scholar. The idiom considered for all related work was only English. The keywords were composed of the terms real-time and ETL and all of the synonyms. The search string was composed of the keywords and all of the synonyms of the terms. So, we used the following basic search string: (“RTDW” OR “Real-Time” OR “Real Time” OR “RealTime” OR “Zero-Latency” OR “Zero Latency”) AND (“ETL” OR “Extract-Transform” OR “Extract Transform” OR “Extract, Transform” OR “Extracting-Transforming” OR “Extracting Transforming” OR “Extracting, Transforming”). It is important to highlight for all sources of search, the string has a slight change in its syntax. So, we made these adjusts on the string to allow us to perform the search into each digital library. Also, we consider finding the keywords on the title or on the abstract of the articles.
12.5
Related Work of Real-Time ETL Process
In this section, we show all related work of real-time ETL process in data warehousing environments. For this purpose, we split the related work into two groups, which are general studies and for applied purposes. Table 12.1 depicts all general studies. The knowledge of all general studies is important to be aware of all theoretical aspects that are involved in the realtime ETL process. It means that we are able to identify the concepts, characteristics and specificities about the ETL process. Moreover, these general studies can aid us to find the state-of-the-art of real-time ETL process, as they expose all content that is available until that time. Despite the fact that the general studies are important to this paper, we will show
94
F. d. A. Vilela and R. R. Ciferri
Table 12.1 Related work which the purpose is general studies Id Author 1 Langseth, J. 2 3
Vassiliadis e Simitsis Penzlin et al.
Title Real-time data warehousing: challenges and solutions Near real time ETL
Current state and future challenges of real-time ETL 4 Farooq e Real-time data warehousing for business Sarwar intelligence 5 Tank et al. Speeding ETL processing in DWs using high-performance joins for changed data capture (CDC) 6 Kakish e Kraft ETL evolution for real-time data warehousing 7 Revathy et al. From DWs to streaming warehouses: a survey on the challenges for real-time data warehousing and available solutions 8 Ferreira et al. Real-time DW: a solution and evaluation 9 Sabry e Ali A survey of real-time DW and ETL 10 Wibowo, A. Problems and available solutions on the stage of extract, transform, and loading in near real-time data warehousing (a literature study) 11 Gajanan Near real-time data warehousing using Mane, N. ETL (extract, transform and load) tools 12 Muddasir e Study of methods to achieve near real Raghuveer time ETL 13 Sabtu et al. The challenges of extract, transform and load (ETL) for data integration in near real-time environment 14 Bouaziz et al. From traditional DW to real time 15 Phanikanth e A big data perspective of current ETL Sudarsan techniques 16 Chandra, H. Analysis of change data capture method in heterogeneous data sources to support RTDW 17 Muddasir e Study of meta-data enrichment methods Raghuveer to achieve near real time ETL
Table 12.2 Related work which the purpose is applied purposes Year 2004 2009 2009 2009 2010
2012 2013
2013 2014 2015
2015 2017 2017
2017 2017 2018
2020
the related work that focuses on applied purposes. Table 12.2 shows these works. Table 12.2 depicts the related work that focuses on applied purposes, which the purpose can be a method, methodology, strategy, technique or architecture. To better explain all related work of applied purposes, the Table 12.3 depicts all related work that is presented in Table 12.2, making a reference of a work of the Table 12.2 by the Id column. Also, all works are classified by the following features: 1. Which is the approach of the paper. In this way, we consider: Method (M), Methodology (Meth), Application (App), Strategy (Str), Technique (Tech), Architecture (Arch), Optimization (Opt);
Id Author Title 1 Bruckner et al. Striving towards near real-time data integration for DWs 2 Viana et al. A real time data extraction, transformation and loading solution for semi-structured text files 3 Naeem et al. An event-based near real-time data integration architecture 4 Chieu, T. e Real-time performance monitoring for Zeng, L. an enterprise information management system 5 Thomsen et al. RiTE: providing on-demand data for right-time data warehousing 6 Santos e Real-time DW loading methodology Bernardino 7 Shi et al. Study on log-based change data capture and handling mechanism in real-time DW 8 Majeed et al. Efficient data streams processing in the real time DW 9 Javed, M. e Data load distribution by semi real-time Nawaz, A. DW 10 Zuters, J. Near real-time data warehousing with multi-stage trickle and flip 11 Zhou et al. An ETL strategy for real-time DW 12 Santos et al. 24/7 Real-time data warehousing: a tool for continuous actionable knowledge 13 YiChuan e Research of real-time DW storage Yao strategy based on multi-level caches 14 MadeSukar sa Change data capture on OLTP staging et al. area for nearly real time DW base on database trigger 15 Tank, D. Reducing ETL load times by a new data integration approach for real-time business intelligence 16 Jain et al. Refreshing datawarehouse in near real-time 17 Xue et al. Measurement of the energy real-time DW system design and Implementation 18 Jia et al. Research on real time DW architecture 19 Obali et al. A real time DW approach for data processing 20 Halenar, R. Real Time ETL Improvement 21 Freudenrei ch An on-demand ELT architecture for et al. real-time BI 22 Valêncio et al. Real time delta extraction based on triggers to support data warehousing 23 Cuzzocrea et Enhancing traditional data warehousing al. architectures with real-time capabilities 24 Lebdaoui et al. An integration adaptation for real-time data warehousing 25 Mao et al. Dynamic mirror based real-time query contention solution for support big real-time data analysis 26 Li e Mao Real-time data ETL framework for big real-time data analysis
Year 2002 2005
2008 2008
2008 2008 2008
2010 2010 2011 2011 2011 2012 2012
2012
2012 2012 2013 2013 2013 2013 2013 2014 2014 2014
2015
(continued)
12
A Survey of Real-Time ETL Process Applied to Data Warehousing Environments
Table 12.2 (continued) Id Author 27 Guo et al.
Title A new ETL approach based on data virtualization 28 Martins et al. AScale: big/small data ETL and real-time data freshness 29 Figueiras et al. An architecture for big data processing on intelligent transportation systems. An application scenario on highway traffic flows 30 Muddasir and CDC and union based near real time Raghuveer ETL 31 Biswas et al. Efficient incremental loading in ETL processing for real-time data integration 32 Godinho et al. ETL framework for real-time business intelligence over medical imaging repositories 33 Machado et al. DOD-ETL: distributed on-demand ETL for near real-time business intelligence 34 Muddasir and A novel approach to handle huge data Raghuveer for refreshment anomalies in near real-time ETL 35 Thulasiram Real time DW updates through and Ramaiah extraction-transformation-loading process using change data capture method
Year 2015 2016 2017
2017 2019 2019
2019 2020
2020
2. The work makes use of the traditional techniques of the ETL process to deal with the data into the operational environment. So, we consider: Yes or No. 3. Which phases of ETL process is approached by the work. So, we consider: Extraction (E), Transformation (T) and Loading (L); 4. Which CDC technique is applied. So, we consider: Log (Log), Trigger (Trigger), Replication (Rep), Delta Files (DF), Timestamp (TS), Snapshot (Ss) and non-mentioned (NM); 5. Which near real-time technique is applied. So, we consider: Direct Trickle Feed (DTF), Trickle and Flip (TF), Real-Time Data Cache (RTDC), Real-Time Partition (RTP), Partial Extraction (PE), Parallelism (P) and non-mentioned (NM). 6. Which real-time requirement is approached by the work. We consider: Availability (Av), Response Time (RT), Scalability (Sca) and Performance (Per). The Table 12.3 shows the comparison of the related work presented in the Table 12.2. As we can note, the majority of the studies focused on the proposal of an architecture to real-time ETL process. Moreover, all studies make use of traditional techniques of ETL process to deal with the data into the operational environment. In other words, the extraction phase of ETL process makes use of a trigger, log file, delta file, snapshot, timestamp or replication to extract data of interest from operational sources.
95
Table 12.3 Comparison of all work which the purpose is general studies Id 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35
Analyzed characteristics 1 2 3 4 Arq Yes ETL NM Arq Yes ET NM Arq Yes EL Trigger Arq Yes EL Trigger, Ts Arq Yes EL NM Meth Yes L NM M Yes E Log Arq Yes E NM Arq Yes E Ts Str Yes L NM Str Yes E Log Meth Yes L NM Arq Yes EL Trigger Arq Yes ETL Trigger Str Yes E NM Arq Yes EL Log App Yes EL Log Arq Yes EL Log Arq Yes EL Log Str Yes T NM Arq Yes ELT NM Arq Yes E Trigger Meth Yes L NM Meth Yes EL Trigger Arq Yes EL NM Arq Yes EL NM Arq Yes TEL NM Arq Yes ETL Log Meth Yes ET NM Arq Yes EL Log Arq Yes E Ss Arq Yes E NM Arq Yes ETL Log Arq Yes L NM Arq Yes L NM
5 NM NM RTP RTP P RTP NM NM NM TF, RTP NM RTP RTDC, RTP NM NM RTP, PE TF TF, RTP TF, RTP NM RTP PE TF, RTP RTP, EP RTDC, RTP RTDC, RTP NM RTDC RTP DTF NM NM RTP RTDC Rep, Ts
6 RT Sca RT RT RT Per RT Per Per RT RT Per Per RT, Per NM Per RT RT RT NM RT, Per RT RT Per Per Per NM RT, Sca RT RT RT RT Av, RT, Sca RT Per
From the point of view of ETL phases, few work deal with all ETL phases, but the majority of work deals with at least the extraction phase. Moreover, the main used CDC technique is log file and trigger. From the point of view of used real-time techniques, the main used techniques are Real-Time Data Cache and Real-Time Partition. Also, the main real-time requirement considered is Response Time and Performance. Moreover, just one work considers the Availability, Response Time and Scalability requirements. The drawback is that the majority of the related work did not compare to each other, they just cite some studies as a related work or as a motivational example. Moreover, the studies did not validated and tested against another approach.
96
F. d. A. Vilela and R. R. Ciferri
Another drawback is that there are newer studies that do not know the previous studies, and almost all studies did not validate experimentally the proposed approach. From all related work, we can highlight the studies that we consider that represent the state-of-the-art for the realtime ETL process. Langseth, J. [20] and Machado et al. [21] represents the state-of-the-art for the real-time ETL process. Langseth, J. [20] developed a work in 2004 and presented some techniques to allow the ETL process to have near real-time capabilities. These techniques are applied in the majority of work that deal with real-time ETL process and in turn it is cited by almost all work. Machado et al. [21] developed an architecture to perform real-time ETL process in data warehousing environments. For this purpose, they make use of some advanced techniques to deal with data that need to be shared between heterogeneous environments. Such techniques are parallelism, buffer and publish/subscribe systems. These techniques showed great performance gain in comparison with the traditional ETL techniques.
3.
4.
5.
6.
7.
8.
9.
12.6
Conclusion and Future Work
This paper presented a literature review of real-time ETL process applied to data warehousing environments. Unlike the previous surveys, this paper showed all related work that proposes a general study or were developed for an applied purpose. Also, we classify the studies by the approach of the paper, if the work consider the traditional technique of the ETL process, which phases of the ETL process is considered, which CDC technique was used, which realtime technique was used to allow perform ETL process with real-time capabilities and which real-time requirement was guaranteed. As a consequence, we pointed out the state-ofthe-art studies in the literature. From this literature review and the positive points and drawbacks, we can point out that the real-time ETL process in the data warehousing environments has a lot of open opportunities of research. Many authors have just an preliminary idea of solutions and they did not reach the complete resolution of the problem. Moreover, unfortunately the majority of the presented solutions was not tested, validated and compared to each other previous related work, which means that we have a wide range of opening opportunities to develop a series of approaches that will be methodologically correct.
10.
11.
12.
13.
14.
15. 16. 17. 18. 19.
References 1. R. Mukherjee, P. Kar, A comparative review of data warehousing ETL tools with new trends and industry insight, in Proceedings - 7th IEEE International Advanced Computing Conference, IACC 2017 (2017) 2. B. Yadranjiaghdam, N. Pool, N. Tabrizi, A survey on real-time big data analytics: Applications and tools, in Proceedings - 2016 Inter-
20. 21.
national Conference on Computational Science and Computational Intelligence, CSCI 2016 (2017), pp. 404–409 H. Chandra, Analysis of change data capture method in heterogeneous data sources to support RTDW, in 2018 4th International Conference on Computer and Information Sciences (ICCOINS) (2018), pp. 1–6. [Online]. Available: https://ieeexplore. ieee.org/document/8510574/ F. Sabry, E. Ali, A survey of real-time data warehouse and ETL. Int. J. Sci. Eng. Res. 5(7) (2014). [Online]. Available: http://www. ijser.org N.M. Muddasir, K. Raghuveer, Study of methods to achieve near real time ETL, in 2017 International Conference on Current Trends in Computer, Electrical, Electronics and Communication (CTCEEC) (2017), pp. 436–441 R. Kimball, J. Caserta, R. Kimball, J. Caserta, The Data Warehouse ETL Toolkit: Practical Techniques for Extracting, Cleaning Conforming, and Delivering Data (Wiley, Hoboken, 2004) K. Kakish, T.A. Kraft, ETL evolution for real-time data warehousing. Conf. Inf. Syst. Appl. Res. 5(2214) (2012). [Online]. Available: www.aitp-edsig.org A. Wibowo, Problems and available solutions on the stage of extract, transform, and loading in near real-time data warehousing (a literature study), in 2015 International Seminar on Intelligent Technology and Its Applications, ISITIA 2015 - Proceeding (2015), pp. 345–349 T. Jain, S. Rajasree, S. Saluja, Refreshing datawarehouse in near real-time. Int. J. Comput. Appl. 46(18), 975–8887 (2012) M. Mesiti, L. Ferrari, S. Valtolina, G. Licari, G. Galliani, M. Dao, K. Zettsu, StreamLoader: An event-driven ETL system for the on-line processing of heterogeneous sensor data, in Advances in Database Technology - EDBT, vol. 2016 (2016), pp. 628–631 X. Li, Y. Mao, Real-time data ETL framework for big real-time data analysis, in 2015 IEEE International Conference on Information and Automation, ICIA 2015 - In conjunction with 2015 IEEE International Conference on Automation and Logistics, Lijiang (2015), pp. 1289–1294 A. Sabtu, N.F.M. Azmi, N.N.A. Sjarif, S.A. Ismail, O.M. Yusop, H. Sarkan, S. Chuprat, The challenges of extract, transform and loading (ETL) system implementation for near real-time environment, in International Conference on Research and Innovation in Information Systems, ICRIIS (2017), pp. 3–7 K.V. Phanikanth, S.D. Sudarsan, A big data perspective of current ETL techniques, in Proceedings - 2016 3rd International Conference on Advances in Computing, Communication and Engineering, ICACCE 2016 (2017), pp. 330–334 M.N. Muddasir, D. Raghuveer, Study of methods to achieve near real time ETL, in International Conference on Current Trends in Computer, Electrical, Electronics and Communication (ICCTCEEC-2017) (2017), pp. 436–441 S. Bouaziz, A. Nabli, F. Gargouri, From traditional data warehouse to real time, in Intelligent Systems Design and Applications (2017) P. Vassiliadis, A. Simitsis, Near real time ETL, in New Trends in Data Warehousing and Data Analysis, vol. 3 (2009) A. Kabiri, D. Chiadmi, Survey on ETL processes. J. Theor. Appl. Inf. Technol. 54(2), 219–229 (2013) P. Vassiliadis, A survey of extract-transform-load technology, International Journal of Data Warehousing and Mining, 5, pp. 1–27. A. Sabtu, N.F.M. Azmi, N.N.A. Sjarif, S.A. Ismail, O.M. Yusop, H. Sarkan, S. Chuprat, The challenges of extract, transform and load (ETL) for data integration in near real-time environment. J. Theor. Appl. Inf. Technol. 95(22), 6314–6322 (2017) J. Langseth, Real-time data warehousing: challenges and solutions, DSSResources.COM (2004) G.V. Machado, Í. Cunha, A.C. Pereira, L.B. Oliveira, ‘DOD-ETL: distributed on-demand ETL for near real-time business intelligence. J. Int. Serv. Appl. 10(1), 1–15 (2019)
Participatory Modeling: A New Approach to Model Graph-Oriented Databases
13
Luis A. Neumann, Enzo Seraphim, Otávio A. O. Carpinteiro, and Edmilson M. Moreira
Abstract
This article presents a new method for modeling graph databases using the entity-relationship model. We analyze four modeling techniques available in the literature, identify the strengths and weaknesses of each one, and propose a method that minimizes the query path while maintaining the clustering base. We performed practical tests on graph bases resulting from these models. The innovation of the new proposal is the participation of the database designer who, through their knowledge of the application’s business rules, interferes in the most appropriate type of mapping in the final model. Keywords
Algorithms · Data mining · Database modelling · Databases optimization · Database systems · Graph database · Information retrieval · NoSQL · Relational databases · Storage structures
13.1
Introduction
Nowadays, companies have an increasing demand to handle large volumes of data in their business processes. To meet this demand, have been presented new database conceptions, such as the so-called NoSQL (Not Only SQL) databases [1], and their use has been increasing every year, according to the website Db-Engines [2] follow-up. The NoSQL databases count on horizontal scalability to manipulate large volumes L. A. Neumann () · E. Seraphim · O. A. O. Carpinteiro E. M. Moreira Federal University of Itajubá, Itajubá-MG, Brazil e-mail: [email protected]; [email protected]; [email protected]
of data per second. Horizontal scalability is based on a distributed memory architecture, including replication and fragmentation of data in different servers [3]. Among several NoSQL data models, the graph-oriented model has the advantage of bringing the technical domain closer to the business domain, facilitating data modeling [4]. The motivation for this work is that many systems initially adopt the relational model but need to be switched to the graph-oriented model when the demand to manipulate a large volume of data increases. Another issue is that most of the applications deployed in NoSQL databases run parallel to data services in relational databases [5]. In these terms, several authors, such as Bordoloi and Kalita [6], Park et al. [7], Lampoltshammer and Wiegand [8], Erven [9], Lysenko et al. [10], Virgilio et al. [11], Pokorn`y [12], Roy-Hubara et al. [13], and Angles [14], present methods to model data in the form of graphs, from the entityrelationship model or the relational model. Based on the above considerations, this paper presents Participatory Modeling, a method to model graph-oriented databases with the main objective of reducing the number of accesses necessary to prospect data, however, without losing the database clustering. According to the methods presented here, the results of the experiments performed on the databases modeled are presented at the end of the article.
13.2
Related Works
This section presents four methods for graph-oriented databases, from modeling performed with the entityrelationship model or from a relational model scheme to identifying edges and nodes of the graph-oriented model. An important concept regarding graph-oriented databases is that there are only nodes and edges. There is no such thing as a table, grouping registers of the same type and same attributes. Thus, when we state that table (or entity) T1 will
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_13
97
98
originate node V1, we are saying that each record of table T1 will become a new type of node V1. Presented by Park et al. [7], the 3NF Equivalent Graph (3EG), part of the relational model, turns entities of the entity-relationship model into nodes and relationships into edges. In its guidelines, this is similar to the solution usually proposed by developers of the Graph Database Management Systems (GDBMS), such as Neo4j [15]. In short, all tables originated from the entity will be a node, and the foreign key attributes will be edges. Tables originated from cardinality unary or binary relationships 1 : 1 or 1 : n will be edges. The other relationships will give a node with edges among the new node and those corresponding to the table connected by them. In this paper, this method is called M01. In other approaches ([6] and [16]), the authors propose modeling based on Reference Graph. Although the steps are different from the previous method, the final result is identical. The main difference between the two methods is that the first is part of the relational model, and the second is of the databases’ entity-relationship model. Entities are converted into nodes. The unary relationships (cardinality 1 : 1 or 1 : n) are converted into self-referent edges. Finally, cardinality 1 : n binary relationships are transformed into node edges from n side to 1 side. All the other relationships generate a node with edges oriented to the involved entities. In this paper, this method is called M02. Several authors, such as Lampoltshammer and Wiegand [8], Erven [9], and Lysenko et al. [10], have approached data modeling based on the simple graph concept: graph without loops and no more than one edge connecting two nodes. These concepts were combined in the Simple Graph Modeling (identified as method M03). This modeling has two phases. The first phase starts in the relational model, following all the Equivalent Graph Modeling steps. In the second phase, all nodes with more than two attributes are analyzed, and the attributes that are not either their primary key or the attribute identifier are removed from the node. These removed attributes generate a new node, being the attribute of their only property. Finally, an edge between the original node and the new node is created. The Simple Graph Modeling (Graph RDF) produces nodes with fewer attributes (the one considered the primary key and the other that identifies the data). This modeling approximates the graph data representation to the representation method called RDF (Resource Description Framework), which proposes to decompose the information into the subject, predicate, and value [17]. The Model-driven design (in this paper called method M04) was proposed by Virgilio et al. [11] and intended to reduce the number of accesses necessary in a base traversal from the junction of two or more nodes into one. Like the other ones, the process starts from the entity-relationship
L. A. Neumann et al.
model and, in short, regroups entities divided into two or more due to the normalization rules [18]. Based on the analysis of the strengths and weaknesses of the four models presented here, we developed the method proposed in this article.
13.3
Participatory Modeling
In this section, we present the Participatory Modeling. It receives this name because it requires the database designer’s participation, besides the modeling algorithms. The designer interferes in the modeling due to his knowledge of the application’s business rules of the database. The process consists of three phases (Fig. 13.1): – Phase 1: the entity-relationship model (ER) is analyzed, and the entities and relationships are modified with the eventual incorporation of entities by relationships or other entities and with the conversion of relationships with a degree (DR ) higher than two into relationships with degree 2, generating a simplified entity-relationship model (ER2). The flowchart of Figs. 13.2, 13.3, and 13.4 summarizes this phase. – Phase 2: the ER2 is migrated to graphs, converting entities into nodes and relationships into edges. – Phase 3: at the designer’s discretion, this phase is performed with the generated graph. The nodes are analyzed, and the designer can transfer identification attributes to another node to reduce the number of attributes of the original node. The first step of Phase 1 (Fig. 13.2) is analyzing each entity int the model to determine which can be removed and have their attributes incorporated by other entities or even incorporated by their relationship. To be eligible as an entity that can be incorporated, the entity shall: – participate in only one relationship; – have a low maintenance profile; that is, once the system is in place, this entity will have few additions of elements and little change from the existing ones. At this moment, the designer’s participation is essential as his analysis will determine the maintenance profile of the entity. This profile represents the Degree of Maintenance of the Entity (DME ), taking the value 1 for low maintenance entities or 0 for high maintenance entities. Entities that meet these two criteria will have Entities’ Incorporation Degree (DEI ) equal to 1; otherwise, DEI will be equal to 0. The process for calculating the incorporation degree is shown in the flowchart of Fig. 13.2.
13 Participatory Modeling: A New Approach to Model Graph-Oriented Databases
99
Fig. 13.1 Proposed method’ flow
Fig. 13.2 Phase 1—step 1—calculation of the entities’ incorporation degree
Phase 1, step 2 (Fig. 13.3), proceeds with analysing each relationship in the model, calculating the Relationship’s Incorporation Degree (DRI ) sequentially and recursively:
– if none of these conditions occur, the result will be the sum of the Entity’s Incorporation Degree (DEI ) of each entity connected by the relationship.
– if the relationship has attributes, then DRI = 0; – else DRI will also be zero (0) case the Relationship Degree (Dn ) is equal to 2, and its cardinality is m : n (m > 1∧n > 1); – otherwise, it will still be zero (0), if Dn = 2, the cardinality is 1 : n, and the entity of n side is DEI = 0;
Then, in step 3 (Fig. 13.4), for each relationship with DRI > 0, that is, which has an entity to be incorporated, our method will perform the following sequence: – if Dn > 2, then remove from the model the entity with DEI = 1, and join its attributes with the relationship attributes;
100
L. A. Neumann et al.
Fig. 13.3 Phase 1—step 2—calculation of the relationship’s incorporation degree I
– otherwise, if Dn = 2, then remove the entity with DEI = 1 from the model, and the respective relationship, incorporating its attributes in the remaining entity; – if two or more entities that participate in the relationship have an incorporation degree equal to 1, one of them shall be elected to be incorporated into the relationship. In this situation, it is up to the designer to define which entity shall be incorporated based on his knowledge. The procedure is performed recursively until the relationship has no more entities that can be removed from the model; that is, all relationships in the model have Relationship’s Incorporation Degree equal to zero (DRI = 0). At the end of this process, relationships of degree greater than 2 (Dn > 2, involving three or more entities) will be replaced: – by the creation of an entity with the same name as the relationship (ER ); – for each entity involved in the relationship (ER ), a relationship of degree 2, cardinality 1 : 1, will be created between the entity ER 1 and ER 2. That done, there are no more relationships of degree greater than 2; all relationships in the model have degree 1 or 2. Once Phase 1 is executed, the new entity-relationship model (ER2) will include the following sets: – original entities of the model (EE ); – entities created from relationships (ER ) (in this phase);
– unary relationships (R1 ); – binary relationships of type 1 : 1 (RR11 ) and 1 : n (RR1n ); – relationships created in this phase (REE ), also binary. The objective of Phase 2 is to map the E-R model resulting from Phase 1 (ER2) for the graph-oriented database model (GDBM). In the first step of Phase 2, for each entity (EE ∧ER ), a node (V) will be created in the MBDG with the same name and with properties, which will be the entity attributes. If the created node does not yet have a primary key property, it will be created for such purpose in the node. The node’s primary key is the property used to identify it on the edges that connect it. After defining the nodes in the first step, the second step identifies and converts the model relationships. Therefore, it will be recognized the unary relationships (RR11 ), binary relationships of cardinality 1 : 1 or 1 : n (RR11 and RR1n ) that were not converted in the previous phase, the binary relationships of cardinality m : n (Rmn ) of the original MER, and the binary relationships 1 : 1 created int the last phase (REE ). Thus: – unary relationships (R1 ), generate bi-directional edges of the node for itself; – binary relationships, RR11 , RR11 , Rmn , and REE , generate bidirectional edges among the nodes corresponding to the entities connected by it; – RR1n relationships generate directed edges of the node corresponding to the n side entity for the node corresponding to the 1 side entity of the relationship. Once this phase is finished, the model is converted.
13 Participatory Modeling: A New Approach to Model Graph-Oriented Databases
101
Fig. 13.4 Phase 1—step 3—calculation of the relationship’s incorporation degree II
13.4
Experiments
The Movies database was used in this work, whose data were obtained in [19] to compare the M01, M02, M03, and M04 methods and validate the method proposed herein. From these methods, we generated a relational database whose entity-relationship model can be seen in Fig. 13.5. We implemented the proposed database in Neo4j, importing data from each entity to the nodes and edges of each method. The first entity from Movies is the Person containing 9505 tuples. A Person has a Name, Date of birth, and Sex. This Person has a performance (ACTED) in the Movie. The performance was in the function (Role) of “Actor,” “Actress,” “Director,” “Screenwriter,” “Producer,” or “Musician” in a Movie, including different interpretations in the same Movie. This entity contains 12,509 tuples. A Movie has a Title, Year of release, and Duration in minutes. The Movie can be classified into different classes (Category), with
a name such as “Drama,” “Comedy,” “Documentary,” etc. Several countries can produce a Movie, but it can only be nominated for the award for only one Country, which has a Name. Annually, movies and persons can be nominated for different cinema awards. Each Award has a Name and different award categories. The different awards can be for the Movie or a Person due to his/her performance in the movie, but the same award cannot be attributed to a movie and a person. For example, the “Best Movie” award is granted only to movies, and the “Best Director” award is granted only to persons who acted in a movie. There are the following relationships between these entities: • degree 2: – IS1, type m : n, between Movie and Category entities; – FROM, type 1 : 1, between Movie and Country entities;
102
L. A. Neumann et al.
Fig. 13.5 Movies—entity-relationship model
• degree 3: – ACTED, among Person, Movie and Role entities; – NOMINATED2, among Movie, Award and TypeAward entities; • degree 4: – NOMINATED1, among Person, Movie, Award and TypeAward entities;
13.4.1 Modeling the Movies Database According to the Participatory Modeling As seen, this method starts with the entity-relationship model, and its first step is to calculate the degree of incorporation of the entities of the model. Table 13.1 shows the calculation of the entities’ incorporation degree. The participatory role of the designer is evidenced: • in the analysis of the Country entity. If the designer considers that it has a low maintenance profile, it can be incorporated into another entity; otherwise, this entity and its relationship will be maintained in the model; • in the analysis of the TypeAward entity, which is related to two different entities, NOMINATED1 and
NOMINATED2, which makes it not incorporable in the first moment. However, according to the designer’s analysis, it can be considered related to only one entity since awards linked to persons are different from awards linked to movies; whoever participates in one relationship does not necessarily participate in another. Based on the entities’ incorporation degree, the relationships’ incorporation degree higher than one is then calculated, the sum of the entities’ incorporation degree that participates in it. Table 13.2 shows this calculation. The ACTED relationship, degree 3, has an incorporation degree equal to 1, resulting from the Role entity. This entity shall be eliminated from the model, and its attributes incorporated by the relationship, which becomes degree 2 and has an incorporation degree equal to 0 (zero). The NOMINATED2 relationship, also degree 3, has an incorporation degree equal to 1, resulting from the TypeAward entity. Therefore, this entity shall be eliminated from the model, and its attributes incorporated by the relationship, which becomes degree 2 and has an incorporation degree equal to 0 (zero). The NOMINATED1 relationship, degree 4, has an incorporation degree equal to 1, resulting from the TypeAward entity. So, this entity shall be eliminated from the model, and its attributes incorporated by the relationship, which becomes degree 3 and has an incorporation degree equal to 0 (zero).
13 Participatory Modeling: A New Approach to Model Graph-Oriented Databases Table 13.1 Movies—method M05—phase 1—entities’s incorporation degree
Entity (EE ) Category Movie Role Award TypeAward Person Country
Number of Relationships it participates in 1 6 1 2 1(*) 3 1
103
Maintenance degree Maintenance low high low high low high low
DME 1 0 1 0 1 0 1
Relationships DRI 1 0 1 0 1 0 1
DEI = if (DME + DRI ) = 2 then 1 else 0 1+1=2→1 0+0=0→0 1+1=2→1 0+0=0→0 1+1=2→1 0+0=0→0 1+1=2→1
(*) Although it has two relationships, records participating in one relationship do not participate in another Table 13.2 Movies—method M05—phase 1—relationships’ incorporation degree
Entities DEI DEI DEI DEI Relationship DR EE1 (EE1 ) EE2 (EE2 ) EE3 (EE3 ) EE4 (EE4 ) IS1 2 Movie 0 Category 1 FROM 2 Movie 0 Country 1 ACTED 3 Movie 0 Person 0 Role 1 NOMINATED1 4 Movie 0 Award 0 TypeAward 1 Person 0 NOMINATED2 3 Movie 0 Award 0 TypeAward 1 (*) has a value of zero because the relationship has a cardinality m : n DRI = DEI (EE1 ) + DEI (EE2 ) + DEI (EE3 ) + DEI (EE4 )
DRI 0(∗) 1 1 1 1
Fig. 13.6 Movies—method M05—end of phase 1—ER2
Fig. 13.7 Movies—method M05—phase 2—converts entities to nodes and relationships to edges
The IS1 relationship, degree 2, is of type m : n, so the relationships’ incorporation degree is equal to 0 and must be left as it is. Finally, the FROM relationship, degree 2, has an incorporation degree equal to 1, resulting from the Country entity, which shall then be incorporated by the Movie entity. There are now only relationships of degree 2 and incorporation degree equal to 0 in the entity-relationship model. Thus, phase 1 is complete This model can be seen in Fig. 13.6. In phase 2, the first step is to convert the entities into nodes and the relationships into edges, as shown in Fig. 13.7. After that, the process is finished, and the graph is generated. Step 3 of the method (the participatory action) provides the possibility of dividing nodes with more than two proper-
ties into two nodes, one with the node’s identifying properties and the second with the other ones. In this database, the attribute Duration of the node Movie could be moved to a second node named OtherMovie, with the same key as the Movie node and this attribute. Likewise, the attribute Birth of the node Person could be moved to the new node (OtherPerson). Therefore, it is up to the database designer to decide whether this step is necessary or not. For this experiment, the nodes were not divided. The final result of the modeling can be seen in Fig. 13.8, where there are also the final models generated by methods M01, M02, M03, and M04, whose databases were created for the comparative experiments of this article.
104
L. A. Neumann et al.
Fig. 13.8 Movies—final result of the five methods
We implemented four databases. We implemented the first according to M01 and M02 methods since the result is the same. To the second database, we implemented it according to the method M03. Regarding the third database, we followed the M04 method. Finally, we implemented the fourth database according to the Participatory Modeling, referred to as the method M05. Ten different queries were performed (Table 13.3) to explore the several generated models’ distinct characteristics. Each query was performed twelve times, discarding the times of the first two queries, for they were considered a warm-up for each database [20]. The queries were sorted according to the Participatory Modeling method’s execution time in an ascending way for comparison with the other methods.
The graphs are presented with the average execution time of each query by method (Fig. 13.9) and the number of accesses to the nodes and edges covered by the query (Fig. 13.10). The Participatory Modeling method was faster than the other methods in almost all queries, except for the query C3 in analyzing each query’s average execution times. In part, the gain in time in the Participatory Modeling method is due to the smaller number of accesses to the nodes and the edges (Fig. 13.10) during the query, compared to other methods. The elimination of nodes and relationships by the proposed method justifies the smaller accesses in the queries since the path in the graph is smaller. However, the result of
13 Participatory Modeling: A New Approach to Model Graph-Oriented Databases
105
Table 13.3 Movies—queries performed 1 2 3 4 5 6 7 8 9 10
Entities Person, movie Person, movie Movie, TypeAward, nominated2 Person, movie, nominated1, TypeAward Person, acting, role, movie Person, movie, nominated1, nominated2, award, category Person, movie, nominated1, nominated2, award, category Movie, country, nominated1, nominated2, award Movie, award, nominated1 Person, acting, movie, nominated1, TypeAward
Fig. 13.9 Movies—average query execution time
Objective Shortest path (Fernanda Torres, Charles Chaplin) All shortest path (Fernanda Torres, Keanu Reeves) Shorter movies that won the oscar Persons with the most awards Persons with the most performances Persons with the most nominations without awards Persons with the most nominations Persons with most acting as director Most nominated films that were awarded in all nominations People with better nominations as best actor/actress
Querie Execution Time (ms)
1.000,0
100,0 M01/M02 M03 M04 M05 10,0
1,0
1
2
3
4
6
5
7
8
9
10
Querie
Fig. 13.10 Movies—nodes crossed by query Nodes + Edges Crossed by Querie
100.000
10.000
1,000 M01/M02 M03
100
M04 M05 10
1
2
query C3, where the path is taken in the proposed method was shorter than in the others, but the time was longer, indicates that we should investigate other factors. Perhaps the results
3
4
5
6
7
8 9 Querie
10
11
12
13
14
15
were affected by the number of attributes or the syntactic structure of the query. We will examine these conjectures in future works.
106
13.5
L. A. Neumann et al.
Conclusion
This article proposes a new modeling method built from the analysis of the best characteristics of four other methods concerning the performance in retrieving information from the graph-oriented database. The main innovation of the proposed method is the conversion of relationships of degrees higher than 2, on which all other methods, like the relational model, create an intermediate node with edges between the involved entities. Besides, in the proposed method, one or more entities can be removed from the model. Its attributes are incorporated as relationship attributes without creating an intermediate node. Thus, the path to be taken when consulting the graph is reduced. Another particularity of the proposed method is the designer’s useful participation, instead of the direct automation through mathematical formulas, in the conversion, since the database designer knows the business data and, based on them, can influence the modeling process. The conducted queries proved that the unification of nodes, with the consequent elimination of edges between them, brings a better performance in queries that return properties that would previously be distributed in different nodes. However, the unification makes performance worse if the query return belongs to the properties of only one of the nodes. What can be concluded is that none of the methods will always be more or less suitable than other ones. Therefore, the method must be chosen according to the application’s characteristics that manipulate the database to be modeled.
References 1. J. Ribeiro, J. Henrique, R. Ribeiro, R. Neto, NoSQL vs. relational database: A comparative study about the generation of the most frequent N-grams, in 2017 4th International Conference on Systems and Informatics (ICSAI) (IEEE, Piscataway, 2017), pp. 1568–1572 2. Kühne-Gasse. DB-Engines Ranking (2020). https://db-engines. com/en/ranking. Accessed 8 Sep 2020
3. R. Cattell, Scalable SQL and NoSQL data stores. SIGMOD Record 39, 12–27, 12 (2010) 4. I. Robinson, J. Webber, E. Eifrem, Graph Databases. (O’Reilly Media, Sebastopol, 2013) 5. F. Ravat, J. Song, O. Teste, C. Trojahn, Efficient querying of multidimensional RDF data with aggregates: Comparing NoSQL, RDF and relational data stores. Int. J. Inf. Manag. 54, 102089 (2020). Elsevier 6. S. Bordoloi, B. Kalita, ER model to an abstract mathematical model for database schema using reference graph. Int. J. Eng. Res. Devel. e-ISSN 6, 51–60 (2013) 7. Y. Park, M. Shankar, B.-H. Park, J. Ghosh, Graph databases for large-scale healthcare systems: A framework for efficient data management and data services, in 2014 IEEE 30th International Conference on Data Engineering Workshops (ICDEW) (IEEE, Piscataway, 2014), pp. 12–19 8. T.J. Lampoltshammer, S. Wiegand, Improving the computational performance of ontology-based classification using graph databases. Remote Sens. 7(7), 9473–9491 (2015) 9. G.C. Galvão Van Erven. MDG-NoSQL: Modelo de Dados para Bancos NoSQL Baseados em Grafos (2015) 10. A. Lysenko, I.A. Roznov˘a¸t, M. Saqi, A. Mazein, C.J. Rawlings, C. Auffray, Representing and querying disease networks using graph databases. BioData Mining 9(1), 23 (2016) 11. R. De Virgilio, A. Maccioni, R. Torlone, Model-driven design of graph databases, in Conceptual Modeling (Springer, Berlin, 2014), pp. 172–185 12. J. Pokorn`y, Conceptual and database modelling of graph databases, in Proceedings of the 20th International Database Engineering & Applications Symposium (ACM, New York, 2016), pp. 370–377 13. N. Roy-Hubara, L. Rokach, B. Shapira, P. Shoval, Modeling graph database schema. IT Professional 19(6), 34–43 (2017) 14. R. Angles, The property graph database model, in AMW (2018) 15. Neo4j (2020). https://neo4j.com/. Accessed 8 Sep 2020 16. S. Bordoloi, B. Kalita, Designing graph database models from existing relational databases. Int. J. Comput. Appl. 74(1), pp. 25–31 (2013) 17. J.F. Sequeda, M. Arenas, D.P. Miranker, On directly mapping relational databases to RDF and OWL, in Proceedings of the 21st international conference on World Wide Web (ACM, New York, 2012), pp. 649–658 18. E.F. Codd, A relational model of data for large shared data banks. Commun. ACM 13(6), 377–387 (1970) 19. Kaggle. Oscar nominations from 1927 to 2015 (2017). https:// www.kaggle.com/theacademy/academy-awards. acessado em 08/setembro/2020 20. D. Gordon, Warm the cache to improve performance from cold start (2015). https://neo4j.com/developer/kb/warm-the-cache-toimprove-performance-from-cold-start/. Accessed 8 Sep 2020
Graph-Based Hierarchical Record Clustering for Unsupervised Entity Resolution
14
Islam Akef Ebeid, John R. Talburt, and Md Abdus Salam Siddique
Abstract
Here we study the problem of matched record clustering in unsupervised entity resolution. We build upon a state-ofthe-art probabilistic framework named the Data Washing Machine (DWM). We introduce a graph-based hierarchical 2-step record clustering method (GDWM) that first identifies large, connected components or, as we call them, soft clusters in the matched record pairs using a graphbased transitive closure. That is followed by breaking down the discovered soft clusters into more precise entity profiles in a hierarchical manner using an adapted graphbased modularity optimization method. Our approach provides several advantages over the original implementation of the DWM, mainly a significant speed-up, increased precision, and overall increased F1 scores. We demonstrate the efficacy of our approach using experiments on multiple synthetic datasets. Our results also provide some evidence of the utility of graph theory-based algorithms despite their sparsity in the literature on unsupervised entity resolution. Keywords
Database · Data mining · Data cleaning · Data quality · Information quality · Graph theory · Graph clustering · Record linkage · Unsupervised entity resolution · Unsupervised data curation
I. A. Ebeid () · J. R. Talburt · M. A. S. Siddique Department of Information Science, University of Arkansas at Little Rock, Little Rock, AR, USA e-mail: [email protected]; [email protected]; [email protected]
14.1
Introduction
Entity resolution is crucial in data cleaning, curation, and integration [1]. It also refers to finding duplicate records within the same table, across various tables, or multiple databases that might refer to the same real-world entities. Traditional and supervised approaches in entity resolution rely on handcrafting rules for matching records. However, the move towards automating entity resolution for data cleaning, curation, and integration has become the goal for many organizations. Thus, unsupervised entity resolution methods have proliferated. However, unsupervised approaches suffer from higher inaccuracies than supervised ones due to inefficiencies, especially in the clustering step, where the goal is to discover unique entity profiles. The base of any entity resolution system is string similarity or fuzzy matching. Traditional and supervised entity resolution relies heavily on human input to guide the entity matching process using predefined rules. Defining those rules depends on handcrafting simple lexical, semantic, and syntactic conditions for matching records based on attribute similarity like the OYSTER system described in [2]. On the other hand, unsupervised entity resolution approaches could be divided into probabilistic, machine learning, and graphbased methods. Some work has been done in probabilistic unsupervised entity resolution that relies on estimating statistical models [3–5]. However, probabilistic methods that rely more on natural language processing, fuzzy matching, and token frequency have proliferated more [6–8]. Probabilistic token-based methods in unsupervised entity resolution typically rely on an automated pipeline of preprocessing, blocking, matching, clustering, and canonicalization. Preprocessing refers to multiple steps that involve
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_14
107
108
merging and parsing data files, tokenizing, and normalizing the unstandardized references. Blocking refers to the strategy used to mitigate the quadratic complexity of pairwise comparisons in unsupervised entity resolution. That strategy relies on quick and dirty techniques that divide the preprocessed unstandardized references into chunks or blocks, avoiding string matching across the whole dataset. Each block can be processed separately, where pairwise string similarity can be applied with less computational cost. The matched pairs from each processed block are then combined and clustered to infer the unique latent entity profiles represented in the data file. The clustering process aims to resolve any conflicts in pairwise matching and find records that indirectly match. Those conflicts typically occur due to the reliance on frequency-based blocking in unsupervised entity resolution systems [1]. In addition, many challenges riddle the unsupervised entity resolution process; the most critical are parameter tuning, finding appropriate thresholds for matching probabilities, and exact record clustering. Also, that process might need to be applied multiple times due to the low accuracy of relying on fuzzy matching alone without rules as in supervised methods.
14.1.1 Contribution Here we study the problem of record clustering in probabilistic unsupervised entity resolution. We propose a graph-based 2-step hierarchical record clustering technique. We modify a state-of-the-art algorithm in unsupervised entity resolution named the Data Washing Machine (DWM) [1]. Integrating both techniques leverage the power and precision of graph theory methods and the speed and robustness of probabilistic token-based methods. Despite its many advantages, the DWM presents with some drawbacks and limitations. For example, the iterative approach used in the DWM relies on assessing the quality of the entity clusters using Shannon’s entropy-based metric. The metric is difficult to interpret due to arbitrary threshold setting; hence it causes unnecessary iterations and higher computational costs. More specifically, we modify the mechanism of iterating over a full, unsupervised entity resolution pipeline multiple times and replace it with a single shot graph-based approach that minimizes and localizes iterations to discover and optimize entity profiles. The optimization approach is done hierarchically and over two steps after remodeling the matched record pairs as edges in a record graph. The first step relies on iteratively discovering large connected components or soft clusters in the unweighted, undirected record graph after applying a threshold to prune the record graph using a graph-based transitive closure algorithm named CC-MR [9]. The second step relies on breaking down those large soft clusters characterized by a high recall into more precise entity
I. A. Ebeid et al.
clusters also iteratively by adapting a graph-based clustering and community detection algorithm called Louvain’s method [10]. The result is a significant speed-up when compared with the original implementation of the DWM. In addition to a significant increase in precision when applied to the synthetic benchmark data set described in [11]. Thus our contributions are summarized as: • Introducing Graph-based Hierarchical Record Clustering (GDWM) by adapting Louvain’s method to a graph-based transitive closure algorithm named CC-MR [9] used in the original DWM. • Contributing to the literature and methods in graph-based unsupervised entity resolution and showing that a hybrid method, combining both probabilistic unsupervised entity resolution pipelines with graph theory approaches, provides better results than using graph-based approaches alone [12] or probabilistic token-based approaches alone [1]. • Performing various experiments on a sampled synthetic benchmark dataset [11] with known ground truths to demonstrate our approach’s effectiveness, accuracy, and robustness. • The code and datasets for our experiments are accessible at https://bitbucket.org/oysterer/dwm-graph/src/master/ GDWM18/ The rest of the paper is organized as follows. In Sect. 14.2, we describe previous work that has been done in graph-based approaches in entity resolution. In Sect. 14.3, we describe both the algorithm and the method proposed. Section 14.4 describes our experiments on the proposed approach using a synthetic dataset in various configurations and settings. In Sect. 14.5, we discuss the results and the implications of the proposed approach. Finally, we conclude in Sect. 14.6.
14.2
Related Work
Here we review the literature on graph-based methods in entity resolution. Despite their sparsity in the literature on unsupervised entity resolution, graph-based methods and algorithms have been adapted before to the problem of entity resolution.
14.2.1 Token-Based Graph Entity Resolution In token-based graph entity resolution, the goal is to construct a bipartite undirected graph of token nodes and record nodes and cluster the record nodes into unique entity profiles using methods such as SimRank [13]. In [14], the authors introduced a graph-based entity resolution model. The model
14 Graph-Based Hierarchical Record Clustering for Unsupervised Entity Resolution
transformed the input data set into a graph of unique tokens where connectivity reflects the co-appearance of tokens in references. The graph was clustered using a weight-based algorithm that considered three types of vertices: exemplar, core, and support vertices. The algorithm then constructed r radius maximal subgraphs from the original token graph to discover clusters related to unique entities. Token-based methods, however, are computationally expensive and memory intensive due to the lack of an integrated blocking strategy.
14.2.2 Record-Record Similarity-Based Graph Entity Resolution Record-record similarity graphs link structured unstandardized references in a weighted undirected graph where the nodes represent unique records. The connectivity represents the degree of similarity between individual references. That approach of constructing a record graph allows to directly utilize a whole set of graph clustering algorithms that graph theory and network science researchers have already developed. While [15] applied a graph clustering algorithm to optimize minimal cliques in the graph by approximating the NPhard graph clique problem through pruning. Moreover, [16] developed the FAMER framework to combine multi-source data using blocking, matching, and clustering schemes. The framework modeled the merged data as a similarity record graph and then leveraged graph clustering techniques to resolve the entities. Other work has leveraged the graph’s structure instead of just the weights between records. In [17], the authors proposed three algorithms to cluster the similarity graph based on structure rather than edge weights. They argue that graph-based transitive closure, such as in [9], produces high recall but low precision because the graph’s structure is not considered during clustering. They justified using maximal clique algorithms to leverage the graph’s structure, which increases precision. There are also centrality and node importance-based methods where the edge weights are not considered, and node scores are propagated, such as in [18]. In addition, the authors introduced the notion of a node resistance score in a co-authorship graph to model entity similarity. Node resistance can be considered a PageRank score [19], where a random walker computes the probability of walking from the source node to the target node iteratively until convergence. Also, in [20], the authors introduced a graph-based model that linked two graph datasets by aggregating similarity scores from neighboring record nodes. Despite their accuracy, record-record similarity methods are complicated and require extensive graph theory knowledge to tune the adapted methods.
109
14.2.3 Hybrid Graph-Based Entity Resolution Hybrid methods that combine token-based bipartite graphs and similarity-based record-record graphs have been investigated in [12] and [21]. The authors proposed an algorithm that combines text similarity with a graph-based algorithm. They first partition the data into a bipartite graph of record pair nodes and frequent term nodes to learn a similarity score of the record pair nodes. Then, the result was used to construct a record-record graph and used to power the CliqueRank algorithm, which runs on the blocks of records identified by the first part, known as the ITER algorithm. The probability of a matching pair of records is then updated iteratively. The authors combined two distinct methods: the random walk-based approach and the graph clustering-based approach. However, the output of their approach is matched pairs of records without introducing any clustering approach that would resolve the entity profiles.
14.3
Method
14.3.1 Framework In this section we describe in detail the proposed framework (GDWM) as shown in Fig. 14.1.
14.3.1.1 Merged Data File Merging data files is a crucial step in entity resolution. Data files usually come in multiple files containing unstandardized duplicate records. Those files are generally of different sizes and layouts, making it challenging for entity resolution algorithms. Merging data files is sometimes done using a single layout and sometimes by combining the files without knowing the specific layout of the metadata. A standard layout means that the data is organized in the same fields. In our case, we merge all incoming data files into one file without relying on a specific layout where each row has a unique identifier, and the unique entity profiles are unknown. More formally, a record set R consists of tuples ri = (ui , fi ) where ui ∈ U is a unique identifier and fi ∈ F is a set of attributes with unknown headers. 14.3.1.2 Parsing The parsing process includes preprocessing and tokenization and is central to our unsupervised entity resolution approach. Tokenization is preceded by a preprocessing step where the text from all the provided attributes and fields is combined into one string and normalized to the upper case. That is followed by removing special characters from the text. We use the standard approach to tokenizing text in English, splitting the text based on spaces between tokens. The parsed
110
I. A. Ebeid et al.
Fig. 14.1 The GDWM framework
data are then loaded into memory, and a dictionary of unique token frequencies is computed. More formally, the processed data file contain a unique distinct set of tokens T with different counts C given that ti is a distinct unique token where ti ∈ T ⊆ fi and ci is the corresponding count of each unique token ti where ci ∈ C and C = |ti ∈ T|.
14.3.1.3 Blocking The goal of the blocking process is to reduce the potential number of pairwise string similarity operations across a data file as much as possible. Blocking usually rely on inexpensive techniques to quickly filter out records in the data file with a very low probability of matching. Frequency-based blocking typically assumes that two records that refer to the same entity should share at least one token [1]. Other blocking techniques include methods such as Locality-Sensitive Hashing [22]. Here we adapt the frequency-based blocking technique presented in the DWM [1]. The blocking method relies on two parameters, beta β, and sigma σ . First each reference ri ∈ R is processed using σ to filter out tokens ti ∈ ri with frequency ci ∈ C and C = |ti ∈ T| above σ . That process could be seen as a stop word removal step where tokens that appear too much are assumed to be nondiscriminative and unimportant. Following for each reference token ti ∈ ri with frequency ci ∈ C and C = |ti ∈ T| below β is considered a blocking token tBi ∈ T. Blocking tokens are identified for each reference in a list L where the filtered records are repeated. The list is then grouped by blocking tokens regardless of reference. Each block will include all the references where the same blocking token appeared, and the number of blocks will be equivalent to the number of unique blocking tokens in the dataset. This process is formalized in Algorithm 2. 14.3.1.4 Pairwise Matching After blocks have been identified, the next step is pairwise matching between references in each block. Even though any string similarity metric can be used on the blocked references to identify matches, we choose to use a string similarity algorithm named the Scoring Matrix used in [1]
Algorithm 1 Blocking algorithm Input : C(T) unique token frequencies, R record set, β , σ Result: B list of blocks B ← list for r ∈ R do if ci ∈ C ∀ tij ∈ r | ti ∈ T ≥ σ then r ← remove tij from r end if ci ∈ C ∀ tij ∈ r | ti ∈ T ≤ β then L ← (tij , r) end end Ls ← sort L, ti ∈ T Tu ← unique tij ∈ Ls for ti ∈ Tu do for ls ∈ Ls do if ti = ti ∈ ls then bi ← r ∈ ls B ← B + bi end end end return B
and developed in [7] and tested in [23] as an adapted MongeElkan [24] algorithm. The Scoring Matrix method relies on lining up stop-word-filtered tokens of both references of the two records at hand as rows and columns of a matrix in order of appearance. The cell values of the matrix are the normalized Levenshtein’s Edit Distance [25] between each token in the first and the second records. After computing the Levenshtein distance between each token, the algorithm then normalizes the edit distance between all the rows and columns in the matrix that are non-zero, dividing by the maximum length between the two tokens. The algorithm then loops on all the rows and columns and averages all the maximum normalized Levenshtein distances in the matrix for each row token or column token, and that becomes the final normalized similarity score between the two references. We will not formalize the Scoring Matrix algorithm here, as more details can be found in [7].
14.3.1.5 Graph-Based Hierarchical Record Clustering (GDWM) As mentioned before, entity resolution aims to find the unique latent entity profiles in a dataset that multiple records might represent. The blocking algorithm considers two similar records having a high probability of representing the
14 Graph-Based Hierarchical Record Clustering for Unsupervised Entity Resolution
111
Fig. 14.2 The typical unsupervised entity resolution approach
same entity if they have a minimum of one token in common. That increases the number of records representing one entity, as shown in Fig. 14.2. As a result, a record may appear in more than one entity, making the process of clustering more expensive and the underlying graph of matched records more complex. In addition, the matching process also implies that one record might be similar in an indirect way to another record by an algebraic property known as transitivity [26]. In a simple sense, if record a matches record b and record b matches record c, then by transitivity record a matches record c. Hence transitive closure is a widely used approach to finding binary relations in ordered sets of pairs. When adapted to entity resolution, it also considers matched records as an edge list in a graph. Hence clustering can be formalized by reformulating the problem as a graph where G = V, E. A vertex/node v ∈ V represents one record r ∈ R where r ≡ v while an edge e ∈ E|E ⊆ V × V represents whether two records ei = (ri , rj ) are matched and an edge weight we ∈ W(E) : E → R represents the normalized similarity between the two records calculated in the pairwise matching step using the Scoring Matrix algorithm [7]. We also define a set of clusters Q ⊆ P(V) where elements of Q are a subset of V and P is a partition of V. The clustered graph conventionally can be seen as a graph of subgraphs where each meta-vertex represents each subgraph of vertices or records. Hence this hierarchical relationship can be formalized as follows in Eqs. (14.1) and (14.2) and also fits in our framework as illustrated in Fig. 14.3: V = Q
(14.1)
E = Qi , Qj : ∃ vi , vj | vi ∈ Qi , vj ∈ Qj , (vi , vj ) ∈ E (14.2)
Here we view the post pairwise matching entity clustering process as a two-step process. We formulate the problem as a graph and apply a hierarchical graph clustering approach to discover the latent entity profiles. Soft clusters are discovered using the efficient graph-based transitive closure algorithm utilized in the DWM [1] and adapted from [9] called CCMR. The input to CC-MR is a pruned unweighted undirected graph of matched pairs. Next, we convert each soft cluster into a complete undirected weighted graph and prune it from the previously computed similarity scores between matched pairs. Finally, a Modularity optimization algorithm adapted from [10] is used to hierarchically discover and refine the original soft clusters increasing the precision of the overall clustering process while eliminating the need for reiterations compared with the DWM framework. Our 2 step hierarchical method is described formally in Algorithm 2. Transitive Closure Using Adapted CC-MR As mentioned before, the ordered list of matched records can be considered an edge list or an adjacency matrix representing a graph of records. Following a simple graph clustering approach to finding the strongly connected components in the graph is used to perform transitive closure as the first step in our 2 step clustering method. CC-MR was introduced in [9] and used in [1] as a single-step clustering method for the DWM. CC-MR starts by creating star subgraphs from the matched pair list. That is done by simply grouping the pair list by the smallest node. Then, the algorithm iteratively checks whether its vertices are assigned to subgraph components where the center of the subgraph component is the smallest vertex in its first-order neighborhood, relying on the MapReduce framework for scalability. The DWM provides an efficient implementation of the algorithm without relying on MapReduce. The algorithm is formalized and described
112
I. A. Ebeid et al.
Fig. 14.3 The matched pairs with their similarity scores are considered an edge list of a graph. Then they are pruned using a threshold μ. Kolb et al. [9] is applied to discover largely connected components or soft
clusters. Blondel et al. [10] is applied to each graph, and final resolved entity clusters are achieved
thoroughly in [9]. Also, we provide a short psuedo-formal description of the implementation we used in Algorithm 4.
the quality threshold is tricky, as shown later in Sect. 14.4.3. In addition, clusters that are good enough but their quality have been slightly below the threshold might stay in the pipeline for 2 or 3 iterations increasing the computational cost unnecessarily. It is worth mentioning that these issues are addressed in a later version of the DWM using non-graphbased approaches. Nevertheless, in the current version that we are working with, those are the issues that motivated us to redesign the entity clustering process allowing us to eliminate the need for a quality metric and reiterations. Furthermore, in our Modularity based hierarchical record graph clustering approach (GDWM), the adapted Modularity-based algorithm is a second step that involves a minimal number of iterations to optimize the Modularity metric over a record-record similarity graph, which does not require threshold setting. We model each soft cluster coming out of the transitive closure process of the DWM as a complete graph. Then, after pruning, the graph of records is further broken down and clustered by optimizing a graph clustering quality heuristic known as Modularity [27]. Modularity is a heuristic often used in graph theory [28] to measure the strength of clusters
Modularity Optimization Using Adapted Louvain Next, we pass the soft clusters produced by the CC-MR transitive closure in addition to the previously computed similarities between records on the block level to our Modularity optimization algorithm adapted from the Louvain method described in [10]. In the original DWM implementation after transitive closure, an evaluation step is applied where each cluster is evaluated using Shannon’s entropy-based algorithm. Then, a normalized quality metric is computed for each cluster, and a threshold epsilon is applied to decide whether the cluster shall be reiterated or canonicalized. The clusters with qualities lower than are reiterated starting from blocking, where the similarity threshold parameter mu μ is incremented, and the lower quality clusters are further broken down. That process carries on until μ reaches one or no clusters left below the quality threshold. Multiple issues emerge with that approach. The most salient is that trying to tune the similarity parameter μ and
14 Graph-Based Hierarchical Record Clustering for Unsupervised Entity Resolution
Algorithm 2 Graph-based hierarchical record clustering Input : S, µ: Pairs of matched records and their computed similarity scores, Mu similarity threshold Result : Q: list of clusters indexed by the least record node in each cluster P ← T ransitiveClosureU singAdaptedCCM R(S, µ) returns connected components as soft clusters see algorithm 3; Ps ← sort(P, r ∈ P ) qc ← unique(r ∈ Ps ) for c ∈ qc do for p(c) ∈ Ps do for n1 ∈ p(c) do for n2 ∈ p(c) do if n1 = n2 and S(n1, n2) exists or S(n2, n1) exists then E ← append(n1, n2, s ∈ S) end end end end V ← initialize node clusters randomly V ← AdaptedLouvain(G(V , E )) see Algorithm 4 V ← append single nodes as separate clusters M ← ComputeM odularity(V , E ) if M ≥ 0 then Q ← rename cluster IDs to least record in cluster V and append end if M ≤ 0 then Q ← assign records in V to one cluster and append end end return Q
Algorithm 3 Transitive closure using adapted CC-MR Input : S, μ: Pairs of matched records and their computed similarity scores, μ similarity threshold Result: P: List of soft clusters as pairs indexed by the least record in the cluster P ← list of soft clusters RP ← initialize list of tuples for s ∈ S do if si ≥ μ then RP ← append(si ) end end RP ← sort by first element in pair while no convergence do for (ri , rj ) ∈ RP do P ← append pair where all record nodes belong to the connected component with the smallest record end end return P
discovered in a graph. More formally, we combine equations (14.1) and (14.2), then reformulate each cluster Q as a separate weighted undirected graph where G = (V, E). Next, the graph is pruned to include only edges available in the dictionary of matched records created in the pairwise matching step, where the edge weights are the precomputed similarity scores. The goal in this second step is to further break down the soft clusters in a hierarchical manner by optimizing through maximizing Modulairty M using the adapted Louvain’s method [10]. The assumption here is that clusters in a graph are often defined by the density of edges between a set of nodes. Hence, the number of connections or edges between clusters characterized by high Modularity is often very low. Modularity is defined in the equation
113
below as M. It measures the difference between the actual number of edges and an expected number of edges between nodes. An expected number of edges between nodes can be considered a random rewiring of the graph given the same nodes. Optimizing Modularity through maximization in a graph is a hard problem and is often tackled through various ways to reduce the number of comparisons between all nodes in a graph. ki kj 1
M= ∂ Qi , Qj Ai,j − 2n i,j 2n
(14.3)
In the equation above, n is the number of edges in the graph, and k is the degree of a node. In addition Ai, j Is a weighted adjacency matrix constructed from E. In addition, i and j are indices for each unique record in the file, represented as a vertex in the graph as v ∈ V. And e ∈ E|E ⊆ V × V and e ≡ Ai,j . Louvain’s method optimizes Modularity by efficiently measuring the difference in Modularities between different configurations of node clustering instead of rewiring the network using a simple equation and procedure [10]. The algorithm starts by randomly assigning each node in the graph to a unique random cluster. Next, a random node is picked and placed in the same cluster as its randomly chosen one of its closest neighbors. Then Modularity difference is measured between when the node was in its cluster and when the node is in the current neighbors’ cluster. If the change in Modularity was positive, then the node is assigned to its new cluster with its neighboring node. If the delta Modularity is negative, the algorithm iterates and places the node in each cluster in the graph until a positive or zero deltas is achieved. If the delta Modularity was zero, the node remains in its cluster. The second phase consolidates the discovered clusters into a new graph of nodes representing whole clusters summing up all involved edge weights. Then the process reiterates on the newly reconstructed graph until no delta Modularity appears. The final clusters from this step can be canonicalized as entity profiles directly. This process is described formally in Algorithm 4.
14.3.1.6 Canonicalization The clusters representing unique entity profiles are defined and persisted to the disk in this final step through a link index file. The link index file describes the final entity clusters of the unsupervised entity resolution framework as a list of ordered pairs where the first element of each pair represents the entity profile identifier the record is assigned to. The entity profile identifier is simply the least record identifier in the cluster. The second element of each pair represents the unique record identifier.
114
I. A. Ebeid et al.
Algorithm 4 Adapted Louvain Input : G = (V , E ): dictionary of nodes with randomly initialized clusters, list of pruned weighted edges as a tuples Result: V : dictionary of nodes reassigned to new clusters while True do Minitial ← ComputeModularity(V , E ) for v1 ∈ V do for v2 ∈ V do update cluster ID of v1 to match cluster ID of v2 ΔM ← compute difference in Modularity M according to Louvain’s reduced equation LM ← (v2 , ΔM) end ci ← arg max (LM ) v2 ← v1 assign the current node to the maximum cluster update V end Mcurrent ← ComputeModularity(V , E ) if Mcurrent ≥ Minitial then G ← whole clusters become new nodes Vreconstructed ,the new edges weights are summation of weights between clusters Ereconstructed end else return V end end
Table 14.1 An excerpt from the synthetic dataset used in the evaluation ID Reference body A985464 LLOYD AARON DEAN 2475 SPICEWOOD DR WINSTON SALEM NC 27106 A956423 LLOYD AARON DEAN 2475 SPICEWOOD DR WINSTON SALEM NC 27106 456 18 2098 A956296 LLOYD AARON D 2475 SPICEWOOD DR WINSTON SALEM NC 27106 456 18 2098 A830349 LLOYD AARON DEAN 2475 SPICEWOOD DR WINSTON SALEM NC 27106 456182098 A875214 ANDREEW AARON STEPHEN 2475 SPICEWOOD DR WINSTON SALEM NC 27106 601 70 6106 A824917 ANDREW AARON STEPHEN 2475 SPICEWOOD DR WINSTON SALEM NC 27106 601 70 6106
Table 14.2 Below is a table describing the statistics of the dataset we are using. A good quality sample indicates that the sample had a limited injected number of errors, while a poor quality sample indicates that a high number of errors were injected
14.4
Experiments
14.4.1 Dataset Our graph-based record clustering method (GDWM) is tested on a synthetic benchmark dataset. We use the synthetic dataset described in [11], a simulator-based data generator that uses probabilistic approaches to generate coherent individual data for persons that do not exist except for S3, which represents generated addresses and names of restaurants that do not exist. The data fields are names, addresses, social security numbers, credit card numbers, and phone numbers mixed in several layout configurations. Some samples are labeled as mixed layout, meaning that each row might come with a different order of those attributes and might not be delimited. The standard label means that all the rows in the data file have the same order and attributes. The generator described in [29] used a probabilistic error model to inject various errors in the previously developed simulated dataset. For example, in this excerpt of a generated data file shown in Table 14.1, the first four records are almost identical except for that record A956296 has a missing last name and the format of the phone numbers or whether they exist at all, all are errors injected and generated on purpose. In addition, the last two records are almost identical except for the first name where an intentional error was introduced. A ground truth set recording the actual clusters of the simulated records is then sampled from the generated synthetic database. Next, the corresponding references are pulled from the generated synthetic database to create various sample files with different sizes, levels of quality, and layouts. Sizes of files can vary from 50 to 20K rows, as shown in Table 14.2.
Sample name S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13 S14 S15 S16 S17 S18
Quality Good Good Moderate Good Good Moderate Good Poor Poor Poor Poor Poor Good Good Good Poor Poor Poor
Layout Standard Standard Standard Standard Standard Standard Mixed Standard Standard Mixed Mixed Mixed Mixed Mixed Mixed Mixed Mixed Mixed
Number of rows 50 100 868 1912 3004 19,998 2912 1000 1000 2000 3999 6000 2000 5000 10,000 2000 5000 10,000
14.4.2 Experimental Setup Here we compare our hierarchical graph-based record clustering approach (GDWM) built using several modules from the DWM with the same version of the original implementation of the DWM. We set the blocking frequency β to 6, and the stop word threshold σ is set to 7 for all runs on all samples. Note that these values are fixed and were set based on experience with running both algorithms. They were picked by observing the execution time of each algorithm and the final F1 score, which will be discussed later. When running both algorithms on each of the 18 samples, we varied μ between 0.1 and 0.9. That gave nine runs per sample.
14 Graph-Based Hierarchical Record Clustering for Unsupervised Entity Resolution Table 14.3 Evaluation metrics and statistics Statistic name True positives
Symbol Description TP Number of record pairs that appeared together in the same cluster correctly True negatives TN Number of record pairs that did not appear in the same cluster correctly False positives FP Number of record pairs that appeared together in the same cluster falsely False negatives FN Number of record pairs that did not appear in the same cluster falsely Precision P TP / (TP + FP) Recall R TP / (TP + FN) F1-score F1 2xP.R/P+R Balanced accuracy A TP(TN + FP) + TN(TP + FN) / (TP + FN)(TN + FP)
When running the DWM, we varied μ and between 0.1 and 0.9. That gave us 81 runs per sample. Both the DWM and our GDWM are implemented in Python with the help of various libraries and packages such as NetworkX [30]. We ran the two setups on an Intel i7 Windows machine with 32 GBs of RAM. The ground truth is a list of each record and its membership cluster identifier. After canonicalization, the saved link index is grouped by the least record identifier in each profile. Thus, all records belonging to the same cluster will have the same record identifier as the first element in the link index pair. We then loop on each pair in the canonicalized link index and examine whether they belong together in the ground truth entity profile. Finally, we measure the following statistics against the ground truth for each sample run (Table 14.3).
14.4.3 Evaluation Here we present the results after running the 18 samples on different similarity thresholds μ; as mentioned before that the DWM ran for each μ level and each level nine times. In addition, the blocking frequency β and the stop word frequency σ are both fixed and set to 6 and 7, respectively. We then chose the value that yielded the best F1 score and lowest execution time in case of a tie. For the GDWM, we ran on each sample 9 times, varying the similarity threshold μ from 0.1 to 0.9. We recorded both runtime statistics. For each sample run for each algorithm, we chose the lowest mu that yielded the highest F1-score, followed by the lowest execution time in case of a tie. We then averaged precision, recall, F1 scores, balanced accuracies, and execution times for both algorithms across the 18 samples. Table 14.4 demonstrates the results given each metric and statistic averaged across all 18 samples. For precision, the improvement appears to be the largest as our approach
115
Table 14.4 Average precision, recall, F1 scores, balanced accuracy, and execution time for each algorithm Balanced Time accu(seconds) Precision Recall F1-score racy 0.7243 0.5782 0.6297 0.7689 203.1
Method DWM (original) GDWM (in 0.8478 0.6862 0.71479 0.8431 24.8 this paper) Bold values indicate that the results from running the proposed framework (GDWM) were better than the results from running the original implementation (DWM)
targets an increased precision relying on the fact that the first step of discovering large soft clusters using the graph-based transitive closure approach CC-MR plays the role of a staging step for the Modularity optimization algorithm through stabilizing the recall even at high mu levels. The next step, which involves Modularity optimization on each soft cluster using Louvain’s method, is responsible for the increased precision. That increased precision can be attributed to first modeling each set of the soft clustered records into a complete graph then weighting the graph using the previously computed similarity scores in the pairwise matching step. That approach automatically prunes the complete graph since not all the records have precomputed similarity scores. That approach also reduces the Modularity optimization done using Louvain and prevents any possible below zero final modularities that often happen when using Louvain’s methods, also known as the resolution limit problem [10]. Second, Louvain’s method is highly efficient when applied to large graphs due to its hierarchical approach in discovering clusters, and it fits nicely as a second step after transitive closure using CC-MR to provide a fully hierarchical graph clustering method. The hierarchical clustering approach also helps preserve the high recall of the entity clusters across the two steps. That is also shown in the increased balanced accuracy, which is a valuable measure for problems such as entity resolution that is usually characterized by having a class imbalance as the number of matched pairs are usually way less than the number of unmatched pairs causing a very high number of true negatives [31]. The significant speed-up achieved across all samples can also be attributed to using the reduced delta Modularity equation provided in the original Louvain’s method paper during optimization [10]. Figure 14.4 shows the interplay between μ, sample size, and F1 scores across the 18 sample run. It is worth mentioning that tuning with the similarity threshold μ can be difficult when running the original implementation of the DWM. Nevertheless, on a sample level and as shown here in Fig. 14.4, our approach, the GDWM, has outperformed the iterative approach used in the DWM, especially at μ levels where GDWM had the most significant overall F1
116
I. A. Ebeid et al.
Fig. 14.4 Best F1 score for each sample with its corresponding Mu level for each algorithm
score. Those μ levels where GDWM outperformed DWM are also higher than the μ levels where DWM had the most significant F1 score on its own. However, at those levels, GDWM outperformance was not as consistent as with the original GDWM μ levels. For example, at μ level 0.3 on poor quality sample files such as S8, S9, S10, S11, S12, S16, S17, and S18, DWM slightly outperformed GDWM. That might be attributed to the existence of a mechanism to iterate and adjust quality levels in the DWM has made it perform better than GDWM in the case of poor data quality files at low levels of μ. From the perspective of parameter sensitivity, at lower μ levels, recall appears to be more sensitive to changes than in higher μ levels. That can be seen in Fig. 14.5, showing 4 sample files with different qualities, sizes, and layouts. The figure demonstrates the sensitivity of the similarity threshold μ with the F1 score, precision, recall, and balanced accuracy. An observation is that in small data samples with good quality like S1 and S2, μ increases precision. Recall seems to be pretty stable across all μ values because it is more sensitive to blocking frequency β and block sizes in general. In large data samples such as S6 with 20K rows, there seems to be an onset for μ where μ before 0.5 does not affect the quality of the clusters produced by GDWM. While low-quality samples like S18 recall starts and continue to be very low at the β we set, affecting the overall performance and F1 scores.
14.5
Discussion
Our GDWM approach eliminates the need to reiterate over a big chunk of the remaining records that have not been well clustered, as is the case in the DWM. Our approach also eliminates the need for using a metric such as Shannon’s entropy that has been utilized in the DWM to evaluate clustering quality. Instead, it uses a Modularity optimization method for graphs through an efficient implementation of Louvain that does not need threshold setting. Combining the graph-based transitive closure approach named CC-MR that has been used in the original implementation of the DWM which acts as a first step in identifying connected components or soft clusters by treating the identified matched pairs of records as an edge list of a graph with a Modularity based clustering approach that we introduced here through the implementation of Louvain’s method, which acted as a second step in identifying and breaking down the original soft clusters, helped in quickly and early on breaking down any soft clusters that existed right after transitive closure hierarchically and efficiently. Our approach, however, is not free of limitations despite its efficacy in identifying latent unique entity profiles. For example, our experiments are highly controlled and applied to a synthetic dataset on data for non-existent individuals;
14 Graph-Based Hierarchical Record Clustering for Unsupervised Entity Resolution
117
Fig. 14.5 Parameter sensitivity on four samples
despite the usefulness of the dataset, it is not diverse enough to explore the interplay between other parameters when applying our approach or any other unsupervised entity resolution approach for that matter. Therefore, we aim to diversify the benchmark datasets used in experimentation in the future. In addition, we plan to study the sensitivity of other crucial parameters like the blocking frequency β and the stop word frequency σ instead of fixing them in experiments. In addition, in the future, we aim at studying how many hierarchical levels of clusters are discovered using our approach and how that can inform the discovery of latent entity profiles using our graph method with fewer optimization iterations.
14.6
Conclusion
This paper demonstrated a graph-based hierarchical record clustering approach to unsupervised entity resolution named GDWM. Our method combines a graph-based transitive closure algorithm with an adapted Modularity-based graph clustering approach based on Louvain’s method in community detection to provide a two-step, efficient hierarchical record clustering capable of producing exact F1 scores in some cases. We built on the work done by Talburt et al. [1], where the authors presented an innovative probabilistic selfassessing and iterative approach to unsupervised entity resolution named the Data Washing Machine (DWM). We integrate our GDWM with the DWM to overcome limitations, especially in self-assessment and reiterations where our ap-
proach eliminated those needs. In addition, we performed various experiments on 18 synthetic benchmark datasets with different sizes, qualities, and layouts to evaluate our adapted algorithm. Finally, we demonstrated the efficacy and advantages our clustering algorithm provides in terms of F1 scores, accuracy, and speed up of execution time compared with the original implementation of the DWM. Acknowledgement This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391.
References 1. J.R. Talburt, A.K., D. Pullen, L. Claassens, R. Wang, An Iterative, self-assessing entity resolution system: first steps toward a data washing machine. Int. J. Adv. Comput. Sci. Appl. 11(12) (2020). https://doi.org/10.14569/IJACSA.2020.0111279 2. J.R. Talburt, Y. Zhou, A practical guide to entity resolution with OYSTER, in Handbook of Data Quality: Research and Practice, ed. by S. Sadiq (Springer, Berlin, 2013), pp. 235–270. https://doi. org/10.1007/978-3-642-36257-6_11 3. T.N. Herzog, F.J. Scheuren, W.E. Winkler, Data Quality and Record Linkage Techniques. Springer Science and Business Media (Springer, New Yrok, 2007) 4. P. Lahiri, M.D. Larsen, Regression analysis with linked data. J. Am. Stat. Assoc. 100(469), 222–230 (2005). https://doi.org/10.1198/ 016214504000001277 5. A. Tancredi, B. Liseo, A hierarchical Bayesian approach to record linkage and population size problems. Ann. Appl. Stat. 5(2B), 1553–1585 (2011). https://doi.org/10.1214/10-AOAS447 6. M. Bilenko, R. Mooney, W. Cohen, P. Ravikumar, S. Fienberg, Adaptive name matching in information integration. IEEE In-
118
7.
8. 9.
10.
11.
12.
13.
14.
15. 16.
17.
18.
I. A. Ebeid et al. tell. Syst. 18(5), 16–23 (2003). https://doi.org/10.1109/MIS.2003. 1234765 X. Li, J.R. Talburt, T. Li, Scoring matrix for unstandardized data in entity resolution, in 2018 International Conference on Computational Science and Computational Intelligence (CSCI) (2018), pp. 1087–1092. https://doi.org/10.1109/CSCI46756.2018.00211 A. Alsarkhi, J.R. Talburt, A method for implementing probabilistic entity resolution. Int. J. Adv. Comput. Sci. Appl. 9(11), 7–15 (2018) L. Kolb, Z. Sehili, E. Rahm, Iterative computation of connected graph components with MapReduce. Datenbank-Spektrum 14(2), 107–117 (2014). https://doi.org/10.1007/s13222-014-0154-1 V.D. Blondel, J.-L. Guillaume, R. Lambiotte, E. Lefebvre, Fast unfolding of communities in large networks. J. Stat. Mech. Theory Exp. 2008(10), P10008 (2008). https://doi.org/10.1088/17425468/2008/10/P10008 J.R. Talburt, Y. Zhou, S.Y. Shivaiah, SOG: a synthetic occupancy generator to support entity resolution instruction and research.. ICIQ 9, 91–105 (2009) D. Zhang, D. Li, L. Guo, K. Tan, Unsupervised entity resolution with blocking and graph algorithms. IEEE Trans. Knowl. Data Eng. 1–1 (2020). https://doi.org/10.1109/TKDE.2020.2991063 G. Jeh, J. Widom, SimRank: a measure of structural-context similarity, in Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2002), pp. 538–543 F. Wang, H. Wang, J. Li, H. Gao, Graph-based reference table construction to facilitate entity matching. J. Syst. Softw. 86(6), 1679–1688 (2013). https://doi.org/10.1016/j.jss.2013.02.026 H. Wang, J. Li, H. Gao, Efficient entity resolution based on subgraph cohesion. Knowl. Inf. Syst. 46(2), 285–314 (2016) A. Saeedi, M. Nentwig, E. Peukert, E. Rahm, Scalable matching and clustering of entities with FAMER. Complex Syst. Inform. Model. Q. 0(16), Art. no. 16 (2018). https://doi.org/10.7250/csimq. 2018-16.04 U. Draisbach, P. Christen, F. Naumann, Transforming pairwise duplicates to entity clusters for high-quality duplicate detection. J. Data Inf. Qual. 12(1), 3:1–3:30 (2019). https://doi.org/10.1145/ 3352591 N. Kang, J.-J. Kim, B.-W. On, I. Lee, A node resistance-based probability model for resolving duplicate named entities. Scientometrics 124(3), 1721–1743 (2020). https://doi.org/10.1007/s11192020-03585-4
19. L. Page, S. Brin, R. Motwani, T. Winograd, The PageRank Citation Ranking: Bringing Order to the Web (Stanford InfoLab, Stanford, 1999) 20. M. Sadiq, S.I. Ali, M.B. Amin, S. Lee, A vertex matcher for entity resolution on graphs, in 2020 14th International Conference on Ubiquitous Information Management and Communication (IMCOM) (2020), pp. 1–4. https://doi.org/10.1109/IMCOM48794. 2020.9001799 21. D. Zhang, L. Guo, X. He, J. Shao, S. Wu, H.T. Shen, A graphtheoretic fusion framework for unsupervised entity resolution, in 2018 IEEE 34th International Conference on Data Engineering (ICDE), Paris (2018), pp. 713–724. https://doi.org/10.1109/ICDE. 2018.00070 22. P. Malhotra, P. Agarwal, G.M. Shroff, Graph-parallel entity resolution using LSH & IMM, in EDBT/ICDT Workshops (2014), pp. 41–49 23. A. Al-Sarkhi, J.R. Talburt, Estimating the parameters for linking unstandardized references with the matrix comparator. J. Inf. Technol. Manag. 10(4), 12–26 (2018) 24. A.E. Monge, C. Elkan, et al., The field matching problem: algorithms and applications, in KDD, vol. 2 (1996), pp. 267–270 25. V.I. Levenshtein, Binary codes capable of correcting deletions, insertions, and reversals, in Soviet Physics Doklady, vol. 10, No. 8 (1966), pp. 707–710 26. S.V. Ovchinnikov, On the transitivity property. Fuzzy Sets Syst. 20(2), 241–243 (1986). https://doi.org/10.1016/01650114(86)90080-1 27. S.E. Schaeffer, Graph clustering. Comput. Sci. Rev. 1(1), 27–64 (2007). https://doi.org/10.1016/j.cosrev.2007.05.001 28. M.E.J. Newman, Modularity and community structure in networks. Proc. Natl. Acad. Sci. U. S. A. 103(23), 8577–8582 (2006). https:// doi.org/10.1073/pnas.0601602103 29. Y. Ye, J.R. Talburt, Generating synthetic data to support entity resolution education and research. J. Comput. Sci. Coll. 34(7), 12– 19 (2019) 30. A. Hagberg, P. Swart, D.S. Chult, Exploring Network Structure, Dynamics, and Function Using NetworkX. Los Alamos National Lab. (LANL) (Los Alamos, NM (United States), 2008) 31. J.P. Mower, PREP-Mt: predictive RNA editor for plant mitochondrial genes. BMC Bioinf. 6(1), 1–15 (2005)
Semantic-MDBScan: An Approach to Assign a Semantic Interpretation to Behavior Changes Detected in Data Stream Scenarios
15
Eldane Vieira Júnior, Rita Maria Silva Julia, and Elaine Ribeiro Faria
Abstract
A great variety of real-world problems can be satisfactorily solved by automatic agents that use adaptive learning techniques conceived to deal with data stream scenarios. The success of such agents depends on their ability to detect changes and on using such information to conveniently adapt their decision-making modules. Several detecting change methods have been proposed, with emphasis on the M-DBScan algorithm, which is the basis of this proposal. However, none of these methods is able to capture the meaning of the identified changes. Thus, the main contribution of this work is to propose an extended version of M-DBScan, called Semantic-MDBScan, with such ability. The proposed approach was validated through artificial datasets representing distinct scenarios. The experiments show that Semantic-MDBScan in fact achieves the intended goal. Keywords
Data stream · Behavior change detection · Semantic assignment · Novelty detection · Classification · Clustering · Markov chain · Spatial entropy · Adaptative learning · M-DBScan
15.1
Introduction
A great variety of problems have been solved by Machine Learning (ML) techniques in different areas as entertainment, financial market, and medicine. Among these techniques, the E. V. Júnior () · R. M. S. Julia · E. R. Faria Computer department, Federal University of Uberlândia, Uberlândia, M.G., Brazil e-mail: [email protected]
batch learning and the adaptive learning can be highlighted [1]. In batch learning, the agents are trained through examples retrieved from databases, being inadequate for dynamic scenarios. Although, adaptive learning can cope with such problems, producing a learning model that is continuously adjusted [2]. Data streams are composed of independent and continuous sequences of events, presenting the following characteristics: the data arrive online; there is no control over data arrival order; there is no limit in the data stream’s length; a processed data sample must be discarded [3, 4]. In non-stationary environments, data distribution is not constant, and this characterizes a phenomenon known as concept drift. Any change perceived in the data distribution of such environments is called a change detection [5]. For agents that act in non-stationary environments, the models are updated according to supervised or unsupervised techniques. Among the supervised ones, the updating can occur whenever there is a decrease in the model performance measure [6], or it can be updated at regular intervals [7]. Concerning both kinds of supervised techniques, it is interesting to point out that the classification of a data sample only expresses the meaning (class) of such sample, not being able to represent the nature of the data samples in a sequence. The unsupervised techniques, similar to the first type of supervised techniques, also update the models. However, differently from those supervised techniques, the unsupervised ones retrieve information from the data themselves (and not from labeled databases). Among them, the Micro-Clustering DBScan (M-DBScan) [8] can be highlighted, since it is the basis of this work. M-DBScan is an unsupervised change detection technique based on entropy and Markov Chain (MC), but not capable to assign a meaning to a change, just like the other supervised or unsupervised data stream technique. In this work, the authors propose an algorithm inspired by M-DBScan—named Semantic-MDBScan—which is able
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_15
119
120
E. V. Júnior et al.
to detect behavior changes in data stream scenarios and to assign a meaning (semantic) to changes. Then, SemanticMDBScan corresponds to an extension of M-DBScan, since it allows the association of an adequate meaning to every data sequence that causes a behavior change. The proposed approach was validated through artificial data sets representing distinct test scenarios. The results show that Semantic-MDBScan succeeded in the tasks of detecting behavior changes in data stream scenarios and of providing meanings to such changes. This paper is organized as following: Sect. 15.2 presents the theoretical foundations; Sect. 15.3 resumes the related works; Sect. 15.4 describes the Semantic-MDBScan; Sect. 15.5 presents the experiments and the results; finally, Sect. 15.6 summarizes the conclusion and points out some future works.
15.2
Theoretical Foundations
This section resumes the main theoretical topics used in this work.
15.2.1 Change Detection in Data Stream A data stream is a sequence of data samples that presents the following characteristics [9]: the data arrives online; there is no control over the order that the data arrives; the data stream size is unlimited; a data already processed is normally discarded, since it deals with limited resources, as memory. In data stream scenarios can occur an event known as concept drift, where the data distribution changes over time [5]. Consider that D represents a data stream, composed by data samples d1 , d2 , d3 , …, which can be organized in sequences S1 , S2 , S3 , …, and each sequence represents a context, composed by samples generated by stationary distributions Gi , where i ≥ 1 [3]. Sequences in succession may present a transition composed of data samples generated by distinct distributions, so a data sample generated by the distribution Gi can be seen as a noise for distribution Gi+1 [3]. Change detection can deal with different types of change: abrupt or gradual. In the abrupt changes, the concept change from one to another suddenly. The gradual changes are more difficult to be identified since the beginning of a change can be seen as noise by the change detection algorithm [3], and the new concept is presented progressively, but over time the new pattern will prevail. Abrupt and gradual changes can present the aspect of a recurrent change, in which, after some time, an already known concept can reappear [5]. The experiments (Sect. 15.5) were elaborated in scenarios of abrupt and gradual change, both with recurrence of concepts.
15.2.2 M-DBScan The algorithm M-DBScan is a technique to detect changes over the unlabeled data obtained from a stream [8]. This algorithm has an incremental clustering process, based on DenStream [10], which is a well-known technique for data stream scenarios. Succeeding the clustering process, a behavior change procedure is applied [11]. The M-DBScan clustering process is composed of two phases, online and offline. The online phase keeps a statistical summary of the data, based on micro-clusters, that maintain information as radius, center, and weight, representing the dynamics of data arrival from the stream. There are different types of micro-clusters, potential-micro-cluster (pmicro-cluster), outlier-micro-cluster (o-micro-cluster), and core-micro-cluster (c-micro-cluster), each of them presents its own characteristics, as described in [10]. The offline phase produces clusters from the result of the online phase [10]. The clusters created in the offline phase are used in a novelty detection module to point out changes over the data. It uses a Markov Chain (MC), and each state in the MC represents a group created in the offline phase. The transitions between the states in the MC are activated considering the dynamics of the data arrivals and the assignment of data to micro-clusters.
15.2.2.1 Novelty Detection The M-DBScan algorithm uses entropy to identify dissimilarities that could be seen as a novelty. Related works as [8, 12, 13] have shown better results using the spatial entropy than the temporal entropy, so, it was chosen for this work. The spatial entropy provides the spatial distribution of the data into clusters. The entropy calculations, and the entropy update, follow the specifications presented in [8, 13]. An entropy update will occur every time that a new data sample is assigned to a micro-cluster in a state of the MC. So, the probabilities in the MC transitions will also be updated. If an MC transition is no longer activated, a decay factor will be applied over its probability value. The detection of a novelty requires that the entropy surpasses its threshold. If it occurs, the novelty is registered in the sliding window to be used later in behavior change detection. 15.2.2.2 Behavior Change Detection The verification of a behavior change happens every time that a novelty is identified. If a behavior change has occurred, it means that the data have presented sufficient variations represented by the novelties in the sliding window. A behavior change detection demands the occurrence of a minimum amount of novelties in an interval of time represented by a sliding window with size k. Every novelty detected is registered in the sliding window, and a novelty can
15
Semantic-MDBScan: An Approach to Assign a Semantic Interpret. to BC in DS Scenarios
be removed from the window if it has slided k times, using as reference the moment that this novelty was registered. Although, if a behavior change were detected, from this point on any novelty will be ignored for a period equal to k. This procedure is applied to prevent the detection of novelties related to an already known change [8].
Finally, it is interesting to point out that all these related works conceived to perform change detection are not able to identify the meaning of a change, which is the main objective of the present work.
15.4 15.3
Related Works
This section presents the state of art in the domain of analyzing data tendency in data stream scenarios according to two distinct purposes: data classification and change detection. Concerning the data classification, the following works can be highlighted: [14–16] and [17]. In these works, the classification only represents the meaning of the sample classified, not being able to portray the sequence. Concerning the change detection objective, according to [3], change detection techniques can be classified into two main categories: (1) the first one monitors changes through performance indicators and (2) the second category monitors the changes by analyzing data distribution on time windows. Next, some works that fit the first category: Drift Detection Method (DDM) [4] deals with scenarios with abrupt alteration in the data, and it assumes that there is a change whenever the classification error increases. The Early Drift Detection Method (EDDM) [4] corresponds to an extension of DDM and it deals with gradual drifts. The work described in [18] introduces the Cumulative Sum Approach (CUSUM), which points out a change whenever a parameter as the classification error exceeds a limit. Page-Hinkley [19] is similar to CUSUM, but it uses the average of a Gaussian signal to define a change. In the second category, there is the technique Adwin [20], which uses two windows that keep information from different moments, and they are statistically compared in order to indicate a change. M-DBScan proposed in [8] also belongs to the second category. In [13], the authors evaluate the performance of the algorithm M-DBScan in a real-time strategy game, which is a scenario that demands effectiveness in the task of detecting change. In [12], the authors used the M-DBScan to show that an adequate representation of the problem can improve the performance of this algorithm. The work in [21] does not fit in any mentioned categories, but it uses a neural network to detect change regions in images and identify their semantic labels. This work does not deal with a non-stationary scenario and according to the authors each period image corresponds to a semantic label, which is different from the semantic of a change in a data sequence.
121
Semantic-MDBScan
The main contribution of this paper is to propose the Semantic-MDBScan: an algorithm that operates in data stream scenarios, able to detect changes in the data sequences, as well as to identify the semantic of the change. For this, Semantic-MDBScan performs a behavior change detection process and, whenever it detects a change, it tries to assign a meaning to it. The detection of behavior changes in Semantic-MDBScan is similar to M-DBScan’s, in which it demands the occurrence of a minimum number of novelties in the sliding window [8]. If Semantic-MDBScan detects a change processing a data sample, it will be classified and the result is used as the semantic of this change. In the flowchart (Fig. 15.1), the dashed-line square delimits the original proceedings of M-DBScan. The algorithm firstly checks whether there is an available data sample in the stream (arrow #1): if so, the algorithm retrieves such sample (arrow #2) and presents it as an input to the online part of the algorithm (arrow #3). In sequence, a decay factor is applied to all micro-clusters but the one to which the last data sample was assigned (arrow #4). The decay factor is used to decrease the relevance of the attributes used to represent these micro-clusters, so they tend to disappear if they do not meet the minimum requirements of a micro-cluster. Then, the algorithm checks whether it is time to perform a rating verification, (arrow #5). If so, the algorithm converts into an o-micro-cluster every p-micro-cluster that does not fulfill anymore the constraints to be considered a p-micro-cluster [10] (arrow #6). Next, the algorithm discards all o-microclusters that do not satisfy anymore the constraints of an omicro-cluster (arrow #7). Then, the offline part of the algorithm (see Sect. 15.2.2) is initialized (arrow #8), but it can be initialized if it is not time to run the rating verification (arrow #9). Once the offline part of the algorithm is concluded, the entropy value and its threshold are calculated (arrow #10 and #11, respectively). At this point, the algorithm has to check whether or not there has already been any behavior change in the processing of the current sliding window (arrow #12). If not (arrow #13), Semantic-MDBScan checks whether the entropy value is greater than its threshold. If it is (arrow #14), it characterizes a novelty occurrence, which is registered in the sliding window, if it is not (arrow #19) it will try to get a new data sample from the stream. In the following, the system checks whether the number of novelties, registered in
122
E. V. Júnior et al.
Fig. 15.1 Semantic-MDBScan flowchart
the sliding window, has reached the minimum value required to point out a change detection (arrow #15). If so (arrow #16), the current data sample is submitted to the Classification module and the label is used as the semantic (meaning) of the new behavior. Next, the algorithm tries to get a new data sample from the stream (arrows #17), this will also happen if no behavior change is detected (arrows #18). If there is a new data sample (arrow #2), the whole process is repeated. Otherwise, the processing of the Semantic-MDBScan is completed (arrow #21). It is important to mention, that if a behavior change is detected, the arrows #13, #14, #15, and #16 will not be executed, for a period equal to the size of the sliding window.
Multilayer Perceptron (MLP) [23]. Both approaches are evaluated through seven distinct datasets. For this, seven labeled datasets—called TrainingDataSet—were created with the purpose of being used by KNN and also in the training of the MLPs of each dataset. All TrainingDataSet and datasets were created from data generated by the framework Massive Online Analysis (MOA) [24], using the random radial basis function generator. The first five datasets present abrupt/recurrent changes, whereas the sixth and seventh ones present gradual/recurrent changes.
15.5.1 Test Datasets
15.5
Experiments and Results
The experiments aim to evaluate the ability of SemanticMDBScan in assigning a correct semantic to behavior changes detected by the algorithm. For comparative purposes, two classification methods will be investigated: the K-Nearest-Neighbor (KNN) [22]; and
In all artificial datasets created for the experiments, real changes occur at fixed interval of m samples, being m = 1000 for Datasets 2, 3, 4 and 6, and m = 20,000 for Datasets 1, 5 and 7. Each sample is represented by 10 attributes. The seven datasets are presented in Table 15.1. In every TrainingDataSet, each behavior (class) counts on a number of samples equal to 15% of m related to the
15
Semantic-MDBScan: An Approach to Assign a Semantic Interpret. to BC in DS Scenarios
15.5.2 Settings in the Experiments
Table 15.1 Description of datasets Dataset
ChangeID
Timestamp
Dataset 1
1 2 3 4 5 6 7 8 9 1 2 3 1 2 3 4 1 2 3 4 1 2 3 4 5 6 7 8 9 1 2 3 1 2 3
20,000 40,000 60,000 80,000 100,000 120,000 140,000 160,000 180,000 1000 2000 3000 1000 2000 3000 4000 1000 2000 3000 4000 20,000 40,000 60,000 80,000 100,000 120,000 140,000 160,000 180,000 1000 2000 3000 20,000 40,000 60,000
Dataset 2
Dataset 3
Dataset 4
Dataset 5
Dataset 6
Dataset 7
123
Semantic value class2 class3 class4 class1 class2 class3 class4 class2 class4 class2 class1 class2 class3 class2 class1 class3 class2 class3 class4 class1 class2 class3 class4 class1 class2 class3 class4 class2 class4 class2 class1 class2 class2 class3 class2
dataset associated to that TrainingDataSet. This value was empirically defined to assure a classification runtime of KNN that does not compromise the efficiency of the real-time processing. Considering the creation of the datasets, one difference refers to the value of the parameter speed, which defines the movement of kernels throughout the arrival of data from the stream: for Dataset 1, it is equal to 500, and for all other datasets it is equal to 100. The test datasets are described in Table 15.1. The columns of such table can be described as following: ChangeID enumerates the behavior change in each dataset; Timestamp indicates the samples when there is a change; Semantic value represents the semantics that prevails in that data sequence.
The experiments were executed using parameters defined empirically. The clustering parameters are: μ = 10; β = 0.105; = 0.45; λ = 0.05, where μ is the minimum number of points required to be a micro-cluster; is the radius limit for micro-cluster; β is the outlier threshold; and λ is the decay factor. Spatial entropy parameters are: minimum number of novelties for a change = 2; size of sliding window = 20. Further: ηs = 0.05; γ = 0.5;δ = 0.02; θ = 2, where ηt is the weighting factor used to control the intensity of the probability update; γ and δ are the weighting factors used in the threshold calculation; and θ is the number of standard deviations used to define a normal distribution for entropy values. In KNN, the value of K was defined through experimental tests in which values from 1 to 10 were evaluated for the datasets described in Table 15.1. The best results were obtained using k = 5, which presented 89% of correct classifications. In order to reduce the FPs in the experiments with Datasets 6 and 7, where the changes are gradual, the following parameters were altered: firstly, the minimum amount of novelties to declare a behavior change is set to 4; secondly, the decay factor (λ) was reduced to 0.001, to slow down the forgetting mechanism of the micro-clusters, which consequently increases the stability of the MC states, reducing the occurrence of FPs. An MLP was created for each dataset, being each one trained using data retrieved from the TrainingDataSet corresponding to such dataset. The architectures of these MLPs were empirically defined throughout the training: 10 neurons in the input layer; 2 hidden layers (the first with 6 neurons and the second with 3). The quantity of neurons of the output layer depends on the dataset to which the MLP will be applied, varying from 2 to 4, since it must have one output neuron for each class in the dataset (Table 15.1). The activation function of the MLPs is the hyperbolic tangent. The learning rate is 0.1, the training is concluded when the error is shorter than 0.01. Tables 15.2 and 15.3 present the results obtained by the experiments involving all the datasets. Such results are presented using the following measures: True Positive (TP): indicates the number of correct change detections with a correct semantic assignment. The delay allowed for the detection is up to 50% of m; False Positive (FP): refers to the number of incorrect change detections produced; False Negative (FN): it occurs in the following situations: (1) The algorithm detects a real change, but after the maximum delay allowed (referred as FN-D); (2) The algorithm detects a real change, but it assigns an incorrect semantic (referred as FN-S); and (3) The algorithm does not detect a change (referred as FN-B, being a blind FN); True Negative (TN): indicates the number of
124
E. V. Júnior et al.
samples that correctly was indicated as not being a behavior change. F1: indicates the balance of precision and recall.
15.5.3 Results Table 15.2 presents the results of the experiments that evaluate the ability of Semantic-MDBScan with KNN to assign an adequate meaning to the changes, and Table 15.3 shows the results produced by Semantic-MDBScan using MLP. Before analyzing the results, it is interesting to remember that the classifiers (KNN or MLP) have no influence over the change detection process, they are only used to assign a meaning to changes. Concerning datasets 1, 2, 3, 4 and 5, with abrupt/recurrent changes, Tables 15.2 and 15.3 show that both SemanticMDBScan versions (KNN and MLP) presented good results. Still analyzing the results shown in these tables for datasets 1, 2, 3, 4 and 5, for both approaches they presented F1-score values that vary from 0.57 to 0.8, which are very good results. Analyzing the number of TPs in the mentioned Tables, it can be observed that the algorithm detected over 50% of the real changes. Figure 15.2 details the occurrences of TPs, FPs, and FNs mentioned in Tables 15.2 and 15.3. Concerning Dataset 6, which presents gradual/recurrent changes, the Semantic-MDBScan, using KNN or MLP, presented an F1-score equal to 0.4 in both approaches. It is Table 15.2 Results of the experiments: Semantic-MDBScan with KNN Dataset Dataset 1 (9 changes) Dataset 2 (3 changes) Dataset 3 (4 changes) Dataset 4 (4 changes) Dataset 5 (9 changes) Dataset 6 (3 changes) Dataset 7 (3 changes)
Measures TP FP 6 6 2 0 3 1 3 2 7 3 1 1 1 13
FN 3 1 1 1 2 2 2
TN 199,985 3997 4995 4994 199,988 3996 79,984
F1 0.57 0.8 0.75 0.66 0.73 0.4 0.11
Table 15.3 Results of the experiments: Semantic-MDBScan with MLP Dataset Dataset 1 (9 changes) Dataset 2 (3 changes) Dataset 3 (4 changes) Dataset 4 (4 changes) Dataset 5 (9 changes) Dataset 6 (3 changes) Dataset 7 (3 changes)
Measures TP FP 6 6 2 0 3 1 3 2 7 3 1 1 2 13
FN 3 1 1 1 2 2 1
TN 199,985 3997 4995 4994 199,988 3996 79,984
F1 0.57 0.8 0.75 0.66 0.73 0.4 0.22
justified by the fact that this dataset is more defiant than the abrupt/recurrent ones. Considering both approaches and all datasets, there were only three FN-S in the experiments, all of them related to Dataset 7 (see Fig. 15.3b), being two associated to SemanticMDBScan with KNN and one produced by SemanticMDBScan with MLP. Such result indicates success in the task of assigning a correct meaning to changes. Concerning the Dataset 7, the value 0.22 of F1-score obtained by Semantic-MDScan with MLP and 0.11 obtained by Semantic-MDScan with KNN, reflects the higher number of FP occurrences obtained in these experiments. It can be explained by the high value of m in this dataset combined with the variation of semantics in the gradual transition in the changes. The occurrence of TP, FP and FN in the experiments with Datasets 6 and 7 can be observed in details in Fig. 15.3. In order to validate the results, the authors applied the Wilcoxon test (tool KEEL [25]). The null hypothesis used is the following one: Semantic-MDScan with MLP does not present a better result than Semantic-MDScan with KNN in terms of correctly assigning a semantic to the detected behavior changes. The results are R+ equal to 17.5 and R− equal to 10.5, with an asymptotic p-value equal to 0.498, which does not allow for the rejection of the null hypothesis. Even being a simpler classifier method than MLP, and having a computational cost superior to the MLP, KNN presented a good performance, since few samples had to be classified, just those that caused a behavior change (for example, in a dataset with m = 1000, it will be 0.1% of the samples or less). Besides that, the incremental characteristic of KNN, turns it into an good classifier for dynamic scenarios, and it does not require a training process, as the MLP does.
15.6
Conclusions and Future Works
This paper presents a new algorithm based on M-DScan, named Semantic-MDBScan, whose purpose is to identify behavior changes in data stream scenarios and to assign an adequate meaning to them. Two distinct techniques to assign such meaning were investigated: MLP and KNN. The Semantic-MDBScan can be useful to build intelligent agents, since they will be capable to perceive changes and adapt their actions based on changes, turning the Semantic-MDBScan into a solution for real-world problems. The results confirm that Semantic-MDBScan succeeded in almost all test scenarios, just presenting a lower performance in databases with gradual/recurrent changes and a long interval between the changes. The results also demonstrated that KNN is adequate to be used in order to reach the purposes pursued by this work, even being a more simple algorithm.
15
Semantic-MDBScan: An Approach to Assign a Semantic Interpret. to BC in DS Scenarios
Fig. 15.2 Detections in datasets 1(a), 2(b), 3(c), 4(d) and 5(e)—abrupt/recurrent changes
125
126
E. V. Júnior et al.
Fig. 15.3 Detections in datasets 6(a) and 7(b)—gradual/recurrent changes
In future works, the authors will use Semantic-MDBScan in games to perceive changes over the adversary’s behavior, and improve the Semantic-MDBScan by better exploring the incremental aspects of KNN.
References 1. K. Faceli, A.C. Lorena, J. Gama, A.C. Carvalho, et al., Inteligência Artificial: Uma abordagem de aprendizado de máquina (LTC, Rio de Janeiro, Brazil, 2011) 2. A. Bifet, R. Gavaldà, Adaptive learning from evolving data streams, in International Symposium on Intelligent Data Analysis (Springer, Berlin, 2009) 3. J. Gama, Knowledge Discovery from Data Streams (CRC Press, Boca Raton, 2010)
4. B. Krawczyk, L.L. Minku, J. Gama, J. Stefanowski, M. Wo´zniak, Ensemble learning for data stream analysis: a survey. Inf. Fusion 37, 132–156 (2017) 5. J. Gama, I. Žliobait˙e, A. Bifet, M. Pechenizkiy, A. Bouchachia, A survey on concept drift adaptation. ACM Comput. Surv. 46(4), 1– 37 (2014) 6. A. Haque, L. Khan, M. Baron, Sand: Semi-supervised adaptive novel class detection and classification over data stream, in THIRTIETH AAAI Conference on Artificial Intelligence (2016) 7. N. Hatamikhah, M. Barari, M.R. Kangavari, M.A. Keyvanrad, Concept drift detection via improved deep belief network, in Iranian Conference on Electrical Engineering (ICEE) (IEEE, Piscataway, 2018) 8. R.M. Vallim, J.A. Andrade Filho, R.F. De Mello, A.C. De Carvalho, Online behavior change detection in computer games. Expert Syst. Appl. 40(16), 6258–6265 (2013) 9. J. Gama, M.M. Gaber, Learning from Data Streams: Processing Techniques in Sensor Networks (Springer, Berlin, 2007)
15
Semantic-MDBScan: An Approach to Assign a Semantic Interpret. to BC in DS Scenarios
10. F. Cao, M. Estert, W. Qian, A. Zhou, Density-based clustering over an evolving data stream with noise, in Proceedings of the 2006 SIAM International Conference on Data Mining. (SIAM, Philadelphia, 2006), pp. 328–339 11. R.M.M. Vallim, Mineração de fluxos contínuos de dados para jogos de computador, Ph.D. Dissertation, Universidade de São Paulo, 2013 12. E. Vieira, R.M.S. Julia, E.R. de Faria, Adapting the markov chain based algorithm M-DBScan to detect opponents’ strategy changes in the dynamic scenario of a starcraft player agent, in Proceedings of International Conference on Agents and AI (SCITEPRESS, Setúbal, 2020) 13. E. Vieira, R.M.S. Julia, E.R. de Faria, Mining data stream to detect behavior change in a real-time strategy game, in Proceedings of Machine Learning and Data Mining in Pattern Recognition (ibai publishing, Leipzig, Germany, 2019) 14. L. Jalali, H. Oh, R. Moazeni, R. Jain, Human behavior analysis from smartphone data streams, in International Workshop on Human Behavior Understanding (Springer, Berlin, 2016), pp. 68–85 15. T. Fawcett, F. Provost, Activity monitoring: Noticing interesting changes in behavior, in Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (1999) 16. F. Mota, M. Paula, I. Drummond, Combined classification models applied to people personality identification, in ITNG 2021 18th International Conference on Information Technology-New Generations (Springer, Berlin, 2021), pp. 457–462
127
17. J. Read, F. Perez-Cruz, A. Bifet, Deep learning in partially-labeled data streams, in Proceedings of the 30th Annual ACM Symposium on Applied Computing (ACM, New York, 2015), pp. 954–959 18. E.S. Page, Continuous inspection schemes. Biometrika 41(1/2), 100–115 (1954) 19. R. Sebastiao, J. Gama, A study on change detection methods, in Progress in Artificial Intelligence, 14th Portuguese Conference on Artificial Intelligence, EPIA (2009), pp. 12–15 20. A. Bifet, R. Gavalda, Learning from time-changing data with adaptive windowing, in Proceedings of the 2007 SIAM International Conference on Data Mining (SIAM, Philadelphia, 2007), pp. 443– 448 21. S. Xiang, M. Wang, X. Jiang, G. Xie, Z. Zhang, P. Tang, Dualtask semantic change detection for remote sensing images using the generative change field module. Remote Sens. 13, 3336 (2021) 22. S.A. Dudani, The distance-weighted k-nearest-neighbor rule. IEEE Trans. Syst. Man Cybern. SMC-6(4), 325–327 (1976) 23. S. Haykin, Neural Networks and Learning Machines, 3/E (Pearson Education India, London, 2010) 24. A. Bifet, G. Holmes, R. Kirkby, B. Pfahringer, MOA: massive online analysis. J. Mach. Learn. Res. 11, 1601–1604 (2010). [Online]. Available: http://portal.acm.org/citation.cfm?id= 1859903 25. J. Alcalá-Fdez, A. Fernández, J. Luengo, J. Derrac, S. García, L. Sánchez, F. Herrera, KEEL data-mining software tool: data set repository, integration of algorithms and experimental analysis framework. J. Multiple-Valued Logic Soft Comput. 17, 255–287 (2011)
A Study on Usability Assessment of Educational Systems
16
Heber Miranda Floriano, Mario Jino, and Ferrucio de Franco Rosa
Abstract
With the massive use of e-learning and Learning Management Systems (LMS) in the education domain, the development of methods and techniques for evaluating the usability of systems is required. This is a critical and important task, as different user profiles interact with educational systems. For example, teachers and students of different ages and limitations (cognitive or physical) demand user-friendly systems. We present a systematic literature review, describing and comparing works that address the usability and accessibility assessment of elearning and m-learning systems. The study provides a current view of the available methodological resources, in addition to pointing out gaps in the literature. This review is aimed at researchers seeking to improve the interface of educational systems. Keywords
Assessment · Distance learning · E-learning · LMS · Survey · Usability
16.1
Nowadays, with the massive use of information and communication technologies, large-scale teaching methods, such as distance learning systems, present various problems [1]. One of the main issues for many user profiles is usability, which is being neglected by educational institutions [2]. Developing usability assessment methods, which could identify possibilities for improvement in educational systems, is imperative. In this context, revealing the state of the art of usability assessment methods aimed at educational systems is necessary. We present a systematic literature review, describing and comparing works that address the usability and accessibility assessment of e-learning and m-learning systems. The remainder of this article is organized as follows: Sect. 16.2 presents related works, i.e., other literature reviews that emerged from the keywords search; in Sect. 16.3 we describe the methodology used in the systematic review; Sect. 16.4 details the results obtained from searches in the scientific databases. Section 16.5 presents a discussion on the analyzed works; Sect. 16.6 presents our conclusions.
16.2 H. M. Floriano () UNIFACCAMP, Campo Limpo Paulista, SP, Brazil M. Jino FEEC/UNICAMP, Campinas, SP, Brazil e-mail: [email protected] F. de Franco Rosa UNIFACCAMP, Campo Limpo Paulista, SP, Brazil CTI Renato Archer, Campinas, SP, Brazil e-mail: [email protected]
Introduction
Related Work
Three literature reviews were identified in our study. The works were classified according to their objectives and applications. Table 16.1 (Appendix I) presents a summary of related works. A synthetic description of these works is presented below. Herbig et al. [3], present a literature review on the user’s cognitive state when using e-learning platforms. Pereira and Rodrigues [4] present a literature review on elearning. The study provides a critical analysis of e-learning projects, and it was divided into (i) apps from recent online
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_16
129
130
H. M. Floriano et al.
mobile stores, and (ii) standalone operating system apps. According to the authors, mobile learning (m-learning) is an extension of distance education, supported by mobile devices with wireless technologies embedded. Hasani et al. [5] present a literature review on UserCentered Design (UCD). UCD is a method used to develop user interfaces for e-learning systems. According to the authors, the most used methods with UCD are questionnaires, interviews, high fidelity prototyping, and usability testing. Projects that used UCD with higher user involvement in various stages of development produced a design with good usability.
16.3
Review Methodology
We present a systematic literature review based on the guidelines proposed by Kitchenham [6] and on the review presented by de Mendonça et al. [7]. We aim to address the following question: “What is the state-of-the-art in usability assessment of educational systems?”. First, exploratory research was carried out to define the study parameters, such as the search time interval, scientific databases, keywords. The search period comprises the years 2010–2021, due to the technology advances in recent years; distance education has taken gigantic proportions. The language defined was English. Adapting the syntax to each database, the following search string was used: (((“Abstract”:usability) OR (“Document Title”:usability)) AND ((“Abstract”:LMS) OR (“Document Title”:LMS))) (((“Document Title”:usability) AND (“Document Title”:elearning)) OR ((“Document Title”:e-learning))). We considered all articles returned from three well-known computer science databases. In the initial search, we obtained 11,645 papers, with 7716 articles from Springer Link, 69 from IEEE Xplore, and 3860 from ACM Digital Library. The inclusion and exclusion criteria were defined by three researchers (two doctors and one expert in the field of computing), as follows: Inclusion: (I1) Research on the usability of e-learning systems and LMS in general; (I2) Research on the accessibility of e-learning systems; (I3) Research on the usability of mobile learning systems; and (I4) Research on the effectiveness of usability of e-learning systems versus face-to-face teaching. Exclusion: (E1) Articles that were written in languages other than English; (E2) Articles related to the usability of e-learning systems; (E3) Publications that are not scientific works; (E4) Abstracts that lack depth or relevant results; and (E5) Articles with more than eleven years since publication. We submitted the collected papers to the inclusion and exclusion criteria. The evaluation considered the title, abstract, and keywords. Thirty-four articles were categorized as works adhering to the research theme and were evaluated in their
entirety according to the criteria. Two researchers analyzed the selected works and generated the final list. Three works are literature reviews related to the topic and are described in Sect. 16.2.
16.4
Results From the Review
The synthetic analysis of the selected studies is organized as follows: (i) Approaches aimed at assessing the usability of educational systems (Sect. 16.4.1); (ii) Approaches aimed at evaluating the usability of mobile learning systems (Sect. 16.4.2); (iii) Approaches aimed at assessing the accessibility of e-learning systems (Sect. 16.4.3). Table 16.2 (Appendix II) presents a synthesis of the analyzed works.
16.4.1 Approaches Aimed at Assessing the Usability of Educational Systems Ternauciuc and Vasiu [8] propose a test strategy for the elearning platform Campus Virtual (based on Moodle). Trilaksono and Santoso [9] proposed an instructional design aimed at kinesthetic students. According to the authors, the proposed design provided a high rate of acceptability, as well as better user learning. Awad et al. [10] evaluated usability features (and related costs) of a commercial LMS platform (Blackboard). The level of user participation and frequency of resource usage were analyzed. Althobaiti and Mayhew [11] performed a usability assessment of an LMS (Jusur). The authors list important factors to be considered when assessing usability, namely: content, learning and support, visual design, navigation, accessibility, interactivity, self-assessment, learning ability, and motivation. Phongphaew and Jiamsanguanwong [12] evaluated an LMS (MyCourseVille) to identify the main interface problems. A usability assessment method was used, associated with five usability attributes, to assess student and teacher interfaces. The method is aimed at LMS developers and can be used as a guideline for evaluating interfaces. Rasheed Hasan [13] evaluated the possibility of using Facebook as an LMS instead of Moodle. According to the authors, despite Facebook having presented some problems regarding usability, students were satisfied with its use as an LMS. Solomon and Makara [14], seeking to understand the effectiveness of online learning environments as pedagogical tools, applied four principles to understand the nature of interactions between students and instructors, namely: (i) Perceptions of lecturers and students about the manner and quality of interactions; (ii) Interactions and cooperation
16 A Study on Usability Assessment of Educational Systems
between students; (iii) Expectations that are communicated in courses that use LMS, as perceived by instructors and students; and (iv) Manners how the use of LMS tools structure and influence the time of teachers and students inside and outside the classroom. Kataoka et al. [1] present a study on language teaching, proposing the effective use of LMS. According to the authors, with the use of information and communication technologies, large-scale language teaching methods present several problems. R. Desai et al. [15] performed comparative evaluations of LMS tools. According to the authors, LMS represents an advance in learning techniques, providing different approaches and teaching techniques. Valerio and Naranjo-Zeledón [16] have described fundamental usability features of LMSs. The characteristics were obtained through the System Usability Scale (SUS) questionnaire, which was implemented in the Moodle and Sakai LMSs. U. Desai et al. [17] evaluated aspects of learning and collaborative problem solving by students. According to the authors, efficient tools are needed to model and evaluate goal-oriented discussion forums built from active student collaborations. Alyami and Alagab [18] evaluated the impact of different learning strategies in virtual environments regarding learning satisfaction. Al-Omar [19] assessed the usability of an LMS using data mining techniques (sentiment analysis and clustering) and the SUS questionnaire. Alvarado et al. [20] applied usability criteria on web and mobile platforms aiming to help students and teachers to reduce problems that could arise, such as difficulty in tests, communication issues, lack of engagement, among others. Little [21] presents the main concerns and functionality problems in the development of LMS and VLE systems. Olivé et al. [22] presented a software framework aimed at identifying students at risk of dropping out of a course. Data from previous versions of the course were used for training the models. The objective is to facilitate the evaluation and use of prediction models in LMS. Zareravasan and Ashrafi [23] and Soledad Fabito et al. [24] present factors that influence students’ intention to use LMS. Fernando et al. [25] present characteristics that affect the quality of e-learning systems. Mtebe and Kissaka [2] propose heuristics to evaluate LMSs. Penha and Correia [26] performed heuristic evaluations of usability in LMSs. Medina-Flores and MoralesGamboa [27] used general heuristic evaluation criteria to develop an approach for evaluating the usability of LMSs by experts. Lehong et al. [28] propose a set of usability guidelines. Rodriguez Morales et al. [29] present a usability study supported by the inquiry method in conjunction with the questionnaire technique, measuring the following indicators:
131
satisfaction, learning, operation, attractiveness, content, and communication. The best results were obtained in the learning and content indicators; with lower scores, those regarding the operation indicator, specifically accessibility and availability. De Freitas et al. [30] present an approach for assessing the impact of gamification on university education, focusing on how data is presented to students. Ugras and Sener [31] evaluated the usability of an online educational platform, focusing on children’s needs. A qualitative usability test was conducted using a multi-method approach.
16.4.2 Approaches Aimed at Evaluating the Usability of Mobile Learning Systems Kumar and Gounda [32] performed a heuristic evaluation on mobile learning applications (m-learning). According to the authors, the interfaces are not effective because they are generic and do not consider access through mobile devices. The use of heuristics could help detect more usability issues in mobile learning apps. Hasan [33] evaluated usability issues in desktop and mobile interfaces from the students’ perspective. Seventeen usability problems were identified in the interfaces of an LMS (Moodle).
16.4.3 Approaches Aimed at Assessing the Accessibility of E-Learning Systems In the context of this work, we consider the term accessibility as a characteristic of usability (broader concept). Lee and Lee [34] evaluated the usability of educational applications for smartphones from the blind users perspective. According to the authors, even when the accessibility requirements were met, the usability level for the applications was low for people with vision disabilities. Melton and Fenwick [35] have developed a user interface that uses voice commands for learning management (Amazon Alexa for Moddle). According to the authors, the integration of Alexa and Moodle features provided people with disabilities access to LMS information in an adequate and faster manner.
16.5
Discussion
Investigating the factors that influence the students’ intention to continue using a LMS is important. The usefulness is a strong indication of intended use [23]. Social influence, information quality, and good university management practices also influence the intended use [24]. The quality of e-learning
132
H. M. Floriano et al.
systems is influenced by system quality, service quality, information quality, instructor competence, user characteristics, individual impact [25]. What can we do to increase user interest in LMSs? According to Rodriguez Morales et al. [29], we need to implement various design improvements to achieve a higher degree of usability. According to de Freitas et al. [30], a quick feedback to students brings various benefits, such as increased motivation, involvement, satisfaction, and improved performance. The use of LMSs is becoming common; however, institutions have not performed usability tests before adopting these systems [2]. Lehong et al. [28] reported a difficulty of interaction faced by students to access knowledge resources, due to generally working in isolation. According to Ugras and Sener [31], there is a lack of usability studies focused on the children’s needs. The authors argue that in terms of information architecture, we should seek to improve the usability of LMS platforms for children.
16.6
Conclusion
We present a systematic literature review on usability assessment of educational systems (e.g., LMS), including mlearning and accessibility issues. The approach used in the literature review allowed the verification and analysis of usability problems. Thirty-three works were selected from 11,645 articles initially retrieved, following inclusion and exclusion criteria. The articles were evaluated, classified, and organized into tables for effortless consultation and comparison. This work is meant to be useful for researchers seeking to develop methods and techniques for assessing usability in LMSs.
A.1
Appendices
A.1.1 Appendix I
Table A.1 Summary of related work Reference Herbig et al. [3] Pereira and Rodrigues [4] Hasani et al. [5] Our Study
Application domain Contribution A B C 1 2 3 4 X X X X X X X X X X X X
Application Domain: (A) E-learning; (B) M-learning; (C) Accessibility. Contribution: (1) Process; (2) Method; (3) Conceptualization; (4) Literature Review
A.1.2 Appendix II Table A.2 Summary of analyzed work Reference Solomon and Makara [14] Little [21] Pereira and Rodrigues [4] Alyami and Alagab [18] Medina-Flores and Morales-Gamboa [27] Ternauciuc and Vasiu [8] R. Desai et al. [15] Ugras and Sener [31] Mtebe and Kissaka [2] Alvarado et al. [20] Althobaiti and Mayhew [11] Trilaksono and Santoso [9] de Freitas et al. [30] Al-Omar [19] Olivé et al. [22] Phongphaew and Jiamsanguanwong [12] Hasan [33] Kataoka et al. [1] Lee and Lee [34] Penha and Correia [26] Rodriguez Morales et al. [29] Lehong et al. [28] Zareravasan and Ashrafi [23] Melton and Fenwick [35] Awad et al. [10] Kumar and Goundar [32] Fernando et al. [25] Valerio and Naranjo-Zeledón [16] Rasheed Hasan [13] Soledad Fabito et al. [24] U. Desai et al. [17]
Application domain Contribution A B C 1 2 3 4 5 6 7 X X X X
X
X X
X
X
X
X X X
X X
X
X
X X
X X
X
X
X X X X
X X X
X X
X X
X X X X
X
X X
X X
X
X
X X
X X
X X
X X
X
X X
X
Application Domain: (A) E-learning; (B) M-learning; (C) Accessibility. Contribution: (1) Process; (2) Method; (3) Conceptualization; (4) Framework; (5) Prototype; (6) Appliance; (7) Methodology
16 A Study on Usability Assessment of Educational Systems
References 1. Y. Kataoka, A.H. Thamrin, J. Murai, K. Kataoka, Effective use of learning management system for large-scale Japanese language education, in Proceedings of the 10th International Conference on Education Technology and Computers, 2018, pp. 49–56. https:// doi.org/10.1145/3290511.3290564 2. J.S. Mtebe, M.M. Kissaka, Heuristics for evaluating usability of learning management systems in Africa, in 2015 ISTAfrica Conference, 2015, pp. 1–13. https://doi.org/10.1109/ ISTAFRICA.2015.7190521 3. N. Herbig, P. Schuck, A. Krüger, User Acceptance of CognitionAware e-Learning: An Online Survey, 2019. https://doi.org/ 10.1145/3365610.3365615 4. O.R.E. Pereira, J.J.P.C. Rodrigues, Survey and analysis of current mobile learning applications and technologies. ACM Comput. Surv. 46(2), 27:1 (2013). https://doi.org/10.1145/ 2543581.2543594 5. L.M. Hasani, D.I. Sensuse, R. R. Suryono, User-centered design of e-learning user interfaces: A survey of the practices, in 2020 3rd International Conference on Computer and Informatics Engineering (IC2IE), 2020, pp. 1–7. https://doi.org/10.1109/ IC2IE50715.2020.9274623 6. B. Kitchenham, Procedures for performing systematic reviews. Keele UK Keele Univ 33(TR/SE-0401), 28 (2004) 10.1.1.122.3308 7. R.R. de Mendonça, F. de Franco Rosa, R. Bonacin, Um estudo sobre análise, representação e detecção de intenções de criminosos em postagens em mídia social, www/internet 2019, p. 27 8. A. Ternauciuc, R. Vasiu, Testing usability in Moodle: When and How to do it, in 2015 IEEE 13th International Symposium on Intelligent Systems and Informatics (SISY), 2015, pp. 263–268. https://doi.org/10.1109/SISY.2015.7325391 9. K. Trilaksono, H.B. Santoso, Moodle based learning management system development for kinesthetic learning style, in 2017 7th World Engineering Education Forum (WEEF), 2017, pp. 602–606. https://doi.org/10.1109/WEEF.2017.8467180. 10. M. Awad, K. Salameh, E.L. Leiss, Evaluating learning management system usage at a small university, in Proceedings of the 2019 3rd International Conference on Information System and Data Mining, 2019, pp. 98–102. https://doi.org/10.1145/3325917.3325929 11. M.M. Althobaiti, P. Mayhew, Assessing the usability of learning management system: User experience study, in E-Learning, EEducation, and Online Training, 2016, pp. 9–18. 12. N. Phongphaew, A. Jiamsanguanwong, Usability evaluation on learning management system, in Advances in Usability and User Experience, 2018, pp. 39–48 13. L. Rasheed Hasan, Is it possible to use facebook instead of moodle Learning management systems (LMS) to support the learning process?,” in 2020 The 4th International Conference on E-Society, E-Education and E-Technology, 2020, pp. 47–51. https://doi.org/ 10.1145/3421682.3421695 14. T.C. Solomon, K. Makara, How does LMS use affect instructional time?, in Proceedings of the 9th International Conference of the Learning Sciences – Volume 2, 2010, pp. 137–138 15. R. Desai, V.K. Ajay, K. Kumar, Sumangali, OSLMS: Open source softwares to E Learning – A comparative study, in 2015 International Conference on Control, Instrumentation, Communication and Computational Technologies (ICCICCT), 2015, pp. 33–37. https://doi.org/10.1109/ICCICCT.2015.7475244 16. C.L. Valerio, L. Naranjo-Zeledón, Usability of open source LMS platforms in academia: Benchmarking from the active learning approach, in 2020 15th Iberian Conference on Information Systems and Technologies (CISTI), 2020, pp. 1–6. https://doi.org/10.23919/ CISTI49556.2020.9141037
133 17. U. Desai, V. Ramasamy, J. Kiper, Evaluation of student collaboration on canvas LMS using educational data mining techniques, in Proceedings of the 2021 ACM Southeast Conference, 2021, pp. 55–62. https://doi.org/10.1145/3409334.3452042 18. S.M. Alyami, A.M. Alagab, The difference in learning strategies in virtual learning environment and their effect on academic achievement and learning satisfaction for distance teaching training program students, in 2013 Fourth International Conference on eLearning “Best Practices in Management, Design and Development of e-Courses: Standards of Excellence and Creativity,” 2013, pp. 102–112. https://doi.org/10.1109/ECONF.2013.40. 19. K. Al-Omar, Evaluating the usability and learnability of the ‘Blackboard’ LMS using SUS and data mining, in 2018 Second International Conference on Computing Methodologies and Communication (ICCMC), 2018, pp. 386–390. https://doi.org/10.1109/ ICCMC.2018.8488038 20. J.V. Alvarado, A.F. Alfaro, M.C. Rivas, C.G. Rodríguez, Collaborative logical framework: An e-learning assesment tool in.LRN platform, in 2016 XI Latin American Conference on Learning Objects and Technology (LACLO), 2016, pp. 1–9. https://doi.org/ 10.1109/LACLO.2016.7751748 21. B. Little, Concerns with learning-management systems and virtual learning environments. ELearn2010(7) (2010). https://doi.org/ 10.1145/1833513.1837142 22. D.M. Olivé, D.Q. Huynh, M. Reynolds, M. Dougiamas, D. Wiese, A Supervised Learning Framework for Learning Management Systems, 2018. https://doi.org/10.1145/3279996.3280014 23. A. Zareravasan and A. Ashrafi, Influencing factors on students’ continuance intention to use learning management system (LMS), in Proceedings of the 9th International Conference on Information Communication and Management, 2019, pp. 165–169. https:/ /doi.org/10.1145/3357419.3357429 24. B. Soledad Fabito, R. L. Rodriguez, A.O. Trillanes, J.I.G. Lira, D.Z. Estocada, P.M.Q. Sta Ana, Investigating the factors influencing the use of a learning management system (LMS): An extended information system success model (ISSM), in 2020 The 4th International Conference on E-Society, E-Education and E-Technology, 2020, pp. 42–46. https://doi.org/10.1145/3421682.3421687 25. E. Fernando, S. Titan, Meyliana, Factors influence the success of E-learning systems for distance learning at the university, in 2020 International Conference on Information Management and Technology (ICIMTech), 2020, pp. 294–299. https://doi.org/10.1109/ ICIMTech50083.2020.9211163 26. M. Penha, W.F.M. Correia, Usability recommendations for a learning management systems (LMS) – A case study with the LMS of IFPE, in Advances in Usability, User Experience and Assistive Technology, 2019, pp. 451–460 27. R. Medina-Flores, R. Morales-Gamboa, Usability evaluation by experts of a learning management system. IEEE Rev. Iberoam. Tecnol. del Aprendiz. 10(4), 197–203 (2015). https://doi.org/10.1109/ RITA.2015.2486298 28. S. Lehong, J. van Biljon, I. Sanders, Open-distance electronic learning environments: Supervisors’ views on usability, in 2019 Conference on Information Communications Technology and Society (ICTAS), 2019, pp. 1–7. https://doi.org/10.1109/ICTAS.2019.8703605 29. G. Rodriguez Morales, P. Torres-Carrion, J. Pérez, L. Peñafiel, Improving the design of virtual learning environments from a usability study, in Information and Communication Technologies of Ecuador (TIC.EC), 2019, pp. 100–115 30. S. de Freitas et al., How to use gamified dashboards and learning analytics for providing immediate student feedback and performance tracking in higher education, in Proceedings of the 26th International Conference on World Wide Web Companion, 2017, pp. 429–434. https://doi.org/10.1145/3041021.3054175
134 31. T. Ugras and O. Sener, A usability study with children on an online educational platform, in Design, User Experience, and Usability: Interactive Experience Design, 2015, pp. 228–239 32. B.A. Kumar, M.S. Goundar, Usability heuristics for mobile learning applications, www/internet 2019, p. 24. 33. L. Hasan, Usability problems on desktop and mobile interfaces of the moodle learning management system (LMS), inProceedings of
H. M. Floriano et al. the 2018 International Conference on E-Business and Applications, 2018, pp. 69–73. https://doi.org/10.1145/3194188.3194192 34. Y. Lee, J. Lee, A checklist for assessing blind users’ usability of educational smartphone applications. https://doi.org/10.1007/ s10209-017-0585-1, p. 24. 35. M. Melton, J. Fenwick, Alexa skill voice interface for the moodle learning management system. J. Comput. Sci. Coll. 35(4), 26–35 (2019)
Part III Cybersecurity
Gesturing with Smart Wearables: An Alternate Way to User Authentication
17
Khandaker Abir Rahman, Avishek Mukherjee, and Kristina Mullen
Abstract
A method of alternate user authentication that relies on sensory data from a smartwatch has been explored in this paper. This attempt to beef up the authentication security was made by taking the user-defined hand gesture into account while wearing a smartwatch. Eventually, the preset hand gesture would work similar way to the password-based authentication scheme. In our experiment, we recorded the 3D coordinate values measured by the accelerometer and gyroscope over a set of gestures. We experimented with 50 gesture samples comprising of five different gesture patterns and ten repeated samples for each pattern. We developed an Android WearOS smartwatch app for sensor data collection, implemented our method of sensor data processing, and performed a series of experiments to demonstrate the potential of this method to achieve high accuracy. Keywords
User authentication · Smartwatch dynamics · Gesture recognition · Gesture-based authentication · Motion-based authentication · Biometrics · Sensors · Gyroscope · Accelerometer · Wear OS
17.1
Introduction and Background
With the increased popularity of devices containing motion sensors that can log user movement and interaction, various modalities including keystroke dynamics (KD) [1–3], mouse K. A. Rahman () · A. Mukherjee · K. Mullen Department of Computer Science and Information Systems, Saginaw Valley State University, University Center, MI, USA e-mail: [email protected]; [email protected]; [email protected]
dynamics [4, 5], touchscreen interaction [6], and device movement [7] have enabled the viability of using behavioral biometrics for user authentication. In recent years, research in gesture recognition methods, have been an active area of interest among security researchers. Some of the earlier work in this area have focused on various security aspects. More recently, user authentication using gesture-based recognition methods have been one such application that has seen varying levels of success. For instance, in [8], a Leap Motion Controller was used to evaluate the usability of various forms of mid-air gestures for authentication. It was concluded by this research that the easier-to-use gestures have weaker security and that users tend to forget their gestures. With the widespread adoption of smart devices, researchers have used the data from multiple on-device sensors for various security purposes, including one-time authentication and continuous user verification. For instance, the research in [7] explored the movement pattern-based authentication using smartphones and found viability in using these movement patterns to thwart common attacks. These studies have helped pave the way for the concept of using smartwatch gestures to achieve user authentication. Initial studies using smartwatch sensors were not pertaining to user authentication. The work in [9], for example, explored the capability of smartwatch devices to detect arm, hand, and finger gestures correctly. It showed the viability in smartwatch motion sensor data to detect unique motion. Similarly, [10] sought to test the validity of fine-motor gestures. The research concluded that five fine-motor gestures chosen by the authors could be recognized with high accuracy supporting the potential for using these gestures in future gesture-based recognition systems. Following the viability of gesture recognition, [11] found that specific gestures could be segmented from continuous movement data, allowing for the use of these gestures as device controllers. The research efforts for gesture-based user authentication using wearables can be separated into two categories - those
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_17
137
138
that involve predetermined gestures, that is gestures that are decided and set for each user before testing, versus those that use user-created gestures, or gestures that are chosen by the user. The work in [12–14] fall into the former category. Even with high accuracy, this approach suffers from one major limitation: the system generated predetermined gestures would be hard to memorize, be easily forgotten, and therefore, lacks usability. On the other hand, the work in [15–17] fall into the second category, that of user-created gestures. In [15], researchers collected motion sensor data from individuals wearing a smart wrist-worn device while they signed their name to validate and distinguish genuine signatures from falsified signatures. Meanwhile, [16, 17] recorded sensor data from a wrist-worn smart device’s and combine the user’s free range of motion while typing their username and password with keystroke dynamics to aid in the authentication of users. However, none of these approaches in [15–17] use hand gestures as a primary mode of user authentication rather, they work as an auxiliary layer of security and triggers when the primary authentication method (signature, username-password, etc.) have been compromised. There has been some research [18, 19], where smartwatch hand-gestures have been considered as the primary mode for user authentication and is closely related to our work. The work in [18], however, is a preliminary study for an authentication system designed exclusively for the smartwatch itself. To be more specific, the method outlined by the authors do not support using the smartwatch to authenticate users on a connected device (computer login, smartphone unlocking, etc.). In contrast, our method can be used as an authentication scheme for any supported device paired with a smartwatch Similarly, [19] designed a motion gesture authentication scheme using Dynamic Time Warping for authentication. The scope of this research was limited to a small set of users who had created their own gestures for use in their experiment. Therefore, the scope for using other pattern matching methods is considerable. In addition, large-scale data analysis would be required to establish the performance of the system. The smartwatch hand gesture recognition method explored in this paper is a standalone user authentication system. In addition, the method developed is intended to authenticate a user not only on the smartwatch itself but also on a connected device (by unlocking the smartphone screen or logging in to a computer, for example) via Bluetooth or WiFi connectivity. We developed a data collection application for the smartwatch, collected 50 user samples collected and implemented our own signal processing methods to identify different user patterns, and finally showed the efficacy over a range of experiments.
K. A. Rahman et al.
The authentication scheme implemented here is also appealing from a security point of view because it possesses inherent resiliency to spoof attacks as authentication samples can be collected without necessarily looking or interacting with the device or account itself. The rest of this paper is organized as follows. Section 17.2 shows the data collection procedures and characteristics of the data. We describe our data processing and experimental setups in Sect. 17.3. Experimental results are shown in Sect. 17.4, and we conclude in Sect. 17.5.
17.2
Data Collection
We used the Android platform to develop a simple application to run smartwatches on the WearOS operating system. The app was built for devices running Android SDK 30 (Android 8.0) or higher. The app consists of a simple interface that shows the current accelerometer and gyroscope values and a button to start-stop recording. Internally, we used Android’s Sensor Manager class to access the raw data from the accelerometer and gyroscope. The app was designed to capture the raw data as fast as the hardware allows access to both the sensors. For both the accelerometer and gyroscope, the sampling rate was typically around 5Hz. Finally, the measured data for each experiment were stored locally in the smartwatch as comma-separated text files and transferred automatically to a common Wi-Fi connected Windows 10 workstation over the Android Debug Bridge (ADB) tool. This auto-transfer feature of gesture samples enables user authentication in a myriad number of scenarios from user login to computer account to a web account, for instance. Hence, the smartwatch acts more like a physical authentication key. The smartwatch used for the data collection was Fossil Gen 4 Carlyle. This smartwatch runs on Android WearOS and was paired with an Android smartphone. For the experiments, five user-created gesture patterns have been developed. Patterns were randomly chosen, keeping them short and easy to remember. Each of these patterns was taken from a sitting position, with the smartwatch wore on the dominant hand. Table 17.1 provides descriptions for each pattern that was created. For each of these patterns, ten samples were recorded over the course of four days, with three samples taken on day one and day two each and two samples taken on day three and day four each. Over the course of four days, a total of 50 samples were recorded across all patterns. Each sample began with a calibration period of around six seconds, with gesture movement immediately following, and then a resting period of around six seconds after the gesture was completed.
17 Gesturing with Smart Wearables: An Alternate Way to User Authentication
139
Table 17.1 Pattern descriptions
3 4
5
Pattern description Movement from the left shoulder down to the right hip, with two finger snaps at the end Forearm movement with the beginning position referred to as center- center, up, center, left, center, down, center, up, center right, center. Arm movement in the shape of a 5-point star Arm movement with the beginning referred to as start, and the first position referred to as center- start, push arm forward into the center position, up, downwards and to the left, center, right, downwards and to the left, center, create a small circle ending the circle back in the center position, back to start. Arm movement beginning by moving the arm up and back down to the starting position twice. When the arm returns to the starting position for the second time, move the wrist to the left and then to the right, passing over the starting position. Return to the left and repeat the movement. When the wrist reaches the right for the second time again, return to the left, and then move the entire arm upwards in a hair tucking behind the ear motion. Finish by following the motion through back to the starting position.
Fig. 17.1 The outcome of the alignment method to identify the start and end of the movement
Amplitude
Pattern number 1 2
X-axis
0
-0.5 0
0.5
1
1.5
Amplitude
Time (nanosecs)
2 ×10
10
Y-axis
1 0.5 0 -0.5 0
0.5
1
1.5
Time (nanosecs)
2 ×10
10
Amplitude
Z-axis 0 -0.5 -1 0
0.5
1
Time (nanosecs)
17.3
Data Processing and Experimental Setups
17.3.1 Pre-processing of Data As mentioned in the previous section, the movement data was collected over several days to add some randomness to the movement. A by-product of this collection process was that there was a timing offset for most of the signals collected. To compensate for this, a stationary period of six seconds was recorded at the start and end of the recorded movement. When comparing two signals, the stationary period above was used to synchronize the signals before computing the squared error between them. The alignment was done using a threshold heuristic based on standard deviation from the first
1.5
2 ×10
10
three seconds of recorded sensor data on both sensors. The starting position of the movement was identified by setting the threshold at least ten times the threshold found in the first three seconds (i.e., stationary period). A similar strategy was implemented to identify the end of the recorded movement by looking at the last three seconds of sensor data. Finally, the process was repeated across all three axes data reported by each sensor, and the final starting and ending timestamps were determined by looking at the mean values reported by every axes. In addition, the signal amplitude was normalized based on the largest amplitude measured across all axes. An example of the alignment result can be seen in Fig. 17.1, where the amplitude of the signal measured on all three axes on the accelerometer is shown. The vertical barriers indicate the subset of the signal data that is clipped and considered during the pre-processing stage. It can be seen
Fig. 17.2 The visual similarity in recorded accelerometer signals for two samples from the same pattern
K. A. Rahman et al.
Amplitude
140
X-axis
1 0 -1 0
5
10
15
20
15
20
15
20
15
20
15
20
15
20
Amplitude
Index
Y-axis
1 0 -1 0
5
10
Amplitude
Index
Z-axis
1 0 -1 0
5
10
Fig. 17.3 Visual dissimilarity in recorded accelerometer signals for samples from two different pattern
Amplitude
Index
X-axis
1 0 -1 0
5
10
Amplitude
Index
Y-axis
1 0 -1 0
5
10
Amplitude
Index
Z-axis
1 0 -1 0
5
10
Index
that the alignment method described above works very well to identify the start and end times of the movement.
17.3.2 Error Computation To compare the reference and the authentication/test samples, the difference between the two sensor signals was computed. First, the length of signals was considered, and the longer
of the two signals was truncated to match the length of the shorter signal. Then the pair-wise squared error between each corresponding amplitude was computed for all axes, and finally, the total difference reported by the accelerometer sensor for comparison was computed. Likewise, an identical process can be used to represent the total difference based on the gyroscope data. As an example, Figs. 17.2 and 17.3 show the computational error when using the accelerometer data across two
17 Gesturing with Smart Wearables: An Alternate Way to User Authentication
sets of signal pairs. Figure 17.2 shows the aligned signals on each of the three axes when comparing a pair of signals that belong to the same pattern. Naturally, after correct alignment, the signal data follows very closely to one another, with a very low error value. On the other hand, Fig. 17.3 shows the aligned signals when looking at signal data from different patterns.
141
longing to the same pattern, whereas the distribution of error values for inter-pattern comparisons are relatively higher, meaning a high level of dissimilarity successfully detected. It can be concluded that our current pre-processing and computational methods work well to distinguish between patterns belonging to different datasets.
17.4.1 Additional Synchronization
17.4
Experimental Results
Our method was evaluated based on two sets of experimental setups – (1) intra-pattern comparison and (2) inter-pattern comparison. In the case of intra-pattern comparison, for each pattern set, a comparison matrix was computed between sample Ri and sample Rj , where i = samples 1, 2, and j = randomly chosen three samples from samples 3 to 10. In other words, the first and second measurements for each pattern were compared against the remaining measurements within the same pattern set. This resulted in a total of 30 comparisons. In the case of inter-pattern comparisons, to compare the performance when measuring different pattern sets, a total of 20 random comparisons were generated. Each pattern was compared against at least one other different pattern to evaluate the performance of our system. Therefore, a total of 50 comparisons were generated using 50 samples. Figure 17.4 shows the distribution of errors on both the accelerometer and gyroscope across all axes. It can be very clearly seen that the relative frequency of lower errors is much higher when computed on pairs beFig. 17.4 Error rates of the accelerometer (left) and gyroscope (right) over three axes
While our current system already produces very good results, the algorithm can be further optimized by improving the alignment accuracy of the signals. Therefore, as a final step, we used the native alignment algorithm called alignsignals [20] implemented in MATLAB to further synchronize the two signals before comparing them. The alignsignals function uses an autocorrelation algorithm to detect if two signals are time-shifted copies of each other. The idea is that for signals belonging to the same pattern, the function should do a very good job at further aligning the signals, whereas, for signals belonging to different pattern sets, the alignment would fail to find a good synchronization since the signals do not have any correlation with each other. The results of applying the alignment function were very encouraging. Figures 17.5 and 17.6 below shows the improvement alignsignals brings when comparing signals belonging to the same pattern. Figure 17.5 is based on the original alignment, while Fig. 17.6 shows the signal data comparison after both pairs have been additionally pre-processed using the above function. The improvement in synchronization is quite significant in the latter case. At the same time, it can
Fig. 17.5 Signal data comparison after applying the original alignment for the same pattern
K. A. Rahman et al.
Amplitude
142
X-axis
1 0 -1
0
5
10
15
20
15
20
15
20
15
20
15
20
15
20
Amplitude
Index
Y-axis
1 0 -1
0
5
10
Amplitude
Index
Z-axis
1 0 -1
0
5
10
Fig. 17.6 Signal data comparison after applying the native alignment algorithm for the same pattern
Amplitude
Index
X-axis
1 0 -1
0
5
10
Amplitude
Index)
Y-axis
1 0 -1
0
5
10
Amplitude
Index)
Z-axis
1 0 -1
0
5
10
Index)
also be seen from the figures Figs. 17.7 and 17.8 that signals belonging to different pattern sets are mostly unaffected by the alignsignals function, and hence, it preserves the level of dissimilarity at a higher level.
17.4.2 Highlights of the Outcomes The experimental setup was re-evaluated using the additional synchronization method discussed above and summarized in
Fig. 17.9. As expected, there was a noticeable improvement in the overall distribution for the intra-pattern and interpattern comparisons compared to the errors shown in Fig. 17.4. In summary, our current method works very well and can be combined with a threshold to establish a secure authentication scheme. Additional synchronization methods such as those described above improve the overall accuracy at the cost of higher computational complexity. Ideally, the accuracy should improve even further over time, as more
Fig. 17.7 Signal data comparison after applying the original alignment for two different patterns
Amplitude
17 Gesturing with Smart Wearables: An Alternate Way to User Authentication
143
X-axis
1 0 -1 0
5
10
15
20
15
20
15
20
15
20
15
20
15
20
Amplitude
Index
Y-axis
1 0 -1 0
5
10
Amplitude
Index
Z-axis
1 0 -1 0
5
10
Fig. 17.8 Signal data comparison after applying the native alignment algorithm for two different patterns
Amplitude
Index
X-axis
1 0 -1 0
5
10
Amplitude
Index
Y-axis
1 0 -1 0
5
10
Amplitude
Index
Z-axis
1 0 -1 0
5
10
Index
measurements are recorded for each pattern, which can be logged from daily use. Overall, our system performed very well in distinguishing signal patterns from different gestures as well as identifying patterns from the same gesture. The accuracy of the system was measured by computing the sum of the squared error between corresponding timestamps across any two patterns across all axes and sensors. Figure 17.10 below shows a comparison between the combined average error on all axes
when comparing signals belonging to the same (Intra) and different (Inter) patterns. It can be seen that the overall error measured by our system on both sensors remains close to 0.05 in more than 80% of cases when comparing signals within the same pattern. On the other hand, more than 90% of inter-pattern comparisons successfully resulted in a higher combined error of 0.2 (or above) across both sensors. Thus, our system can be used to establish a clear threshold between different patterns, thereby providing a reliable authentication scheme.
144
K. A. Rahman et al.
Fig. 17.9 Error rates of the accelerometer (left) and gyroscope (right) over three axes applying the alignment function
Overall Error Comparison
1
17.5
Fraction
0.8
0.6
0.4 Inter (Accelerometer) Inter (Gyroscope) Intra (Accelerometer) Intra (Gyroscope)
0.2
0 0
0.05
0.1
0.15
0.2
0.25
0.3
Average Error Fig. 17.10 Overall error for inter and intra-pattern comparisons
It is also noticeable that the accelerometer in general, performs better by giving off lower error rates compared to the gyroscope errors for intra-pattern comparisons. On the other hand, gyroscope performs better by giving off higher error rates for inter-pattern comparisons. Likewise, the Y and Z axes overall perform better than X-axis, especially when it comes to inter-pattern comparisons. Therefore, applying more weights on the better-performing sensors and/or axes towards decision making (accept/reject the user) would be an interesting dimension for future exploration.
Conclusion and Future Works
This work explores hand gestures as a mode for user authentication using the smart wearable. We analyzed 50 samples for five hand gesture patterns that had signal data from two sensors and three axes each. We performed a series of inter and intra pattern comparisons to demonstrate the efficacy. Our method achieved an error rate close tov0.05 for around 80% of the intra-pattern comparisons and also successfully generated higher error rates above 0.2 for more than 90% of the inter-pattern comparisons. We believe the scope for improving the accuracy (i.e., decreasing the error for intra-pattern comparison and increasing for inter-pattern comparison) is still at large. Therefore, for future works, we can regard the following for immediate improvements of our methodology – (1) apply weighted fusion methods to combine accelerometer and gyroscope comparisons together, (2) generate higher-level features such as total duration, the distance between peaks, and (3) consider sophisticated pattern matching algorithms such as Principal Control Analysis, Support Vector Machine, and Convolution Neural Network. Acknowledgment Research reported in this publication was supported in part by funding provided by the National Aeronautics and Space Administration (NASA) under award number 80NSSC20M0124.
17 Gesturing with Smart Wearables: An Alternate Way to User Authentication
References 1. D. Gunetti, C. Picardi, Keystroke analysis of free text. ACM Trans. Inf. Syst. Secur. 8(3), 312–347 (2005) 2. K. A. Rahman, D. Neupane, A. Zaiter, M. Hossain, Web user authentication using chosen word keystroke dynamics in 18th IEEE International Conference on Machine Learning and Applications, USA, 2019 3. F. Monrose, A. Rubin, Authentication via keystroke dynamics, in 4th Conference on Computer and Communication Security, USA, 1997 4. K. A. Rahman, R. Moormann, D. Dierich, M. Hossain, Continuous user verification via mouse activities, in 8th International Conference on Multimedia Communication, Services and Security, Poland, 2015 5. C. Shen, Z. Cai, X. Guan, Y. Du, R. Maxion, User authentication through mouse dynamics. IEEE Trans. Inf. Forensics Secur.8(1), 16–30 (2013) 6. K. A. Rahman, J. Maes, How discernible user impromptu behavior is when unlocking touch screen? in International Conference on Computer and Information Technology, Bangladesh, 2017 7. K. A. Rahman, D. Tubbs, M. Hossain, Movement pattern based authentication for smart mobile devices in 17th International Conference on Machine Learning and Applications, 2018, pp. 1054– 1058 8. X. Xu, J. Li, Y. Lian, X. Xiao, S. Qu, X. Wang, Which one to go: Sec. and usability evaluation of mid-air gestures. CoRR abs/1811.10168 (2018) 9. C. Xu, P. Pathak, P. Mohapatra, Finger-writing with smartwatch: A case for finger and hand gesture recognition using smartwatch, in 16th International Workshop on Mobile Computing Systems and Applications, USA, 2015, pp. 9–14 10. H. Wen, J. Rojas, A. Dey, Serendipity: Finger gesture recognition using an off-the-shelf smartwatch, in CHI Conference on Human Factors in Computing Systems, USA, 2016, pp. 3847–3851
145
11. P. Zhu, H. Zhou, S. Cao, P. Yang, S. Xue, Control with gestures: A hand gesture recognition system using off-the-shelf smartwatch, in 4th International Conference on Big Data Computing and Communications, Chicago, IL, USA, 2018, pp. 72–77 12. J. Yang, Y. Li, M. Xie, MotionAuth: Motion-based authentication for wrist worn smart devices, in International Conference on Pervasive Computing and Communication, USA, 2015, pp. 550–555 13. X. Yu, Z. Zhou, M. Xu, X. You, X. Li, ThumbUp: Identification and authentication by smartwatch using simple hand gestures, in International Conference on Pervasive Computing and Communication, USA, 2020, pp. 1–10 14. Y. Zhao, R. Gao, H. Tu, Smartwatch user authentication based on the arm-raising gesture. Interact. Comput. 32(5–6), 569–580 (2020) 15. G. Li, H. Sato, Handwritten signature authentication using smartwatch motion sensors, in 44th Annual Computers, Software, and Applications Conference, Spain, 2020, pp. 1589–1596 16. B. Chang, X. Liu, Y. Li, P. Wang, W.T. Zhu, Employing smartwatch for enhanced password authentication, in Wireless Algorithms, Systems, and Applications, Lecture Notes in Computer Science, vol. 10251, (Springer, 2017) 17. B. Chang, Y. Li, Q. Wang, W.T. Zhu, R.H. Deng, Making a good thing better: Enhancing password/PIN-based user authentication with smartwatch. Cybersecurity 1, 7 (2018) 18. A. Lewis, Y. Li, M. Xie, Real time motion-based authentication for smartwatch, in International Conference on Communication and Network Security, 2016, pp. 380–381 19. Y. Li, M. Xie, Understanding secure and usable gestures for realtime motion based authenticationa, in INFOCOM 2018, pp. 13–20 20. Native Alignment Algorithm in MATLAB. URL: https:// www.mathworks.com/help/signal/ref/alignsignals.html
Software Optimization of Rijndael256 for Modern x86-64 Platforms Nir Drucker
and Shay Gueron
Abstract
The AES was standardized in 2001 by NIST and has become the de facto block cipher used today. AES is a block cipher with a block size of 128 bits and is based on the proposal by Rijmen and Daemen, named “Rijndael”. The Rijndael proposal includes a definition for a block cipher with 256 bits block size (and a 256-bits key), which we call here Rijndael256. This variant has not been standardized. This paper describes software optimization methods for fast computations of Rijndael256 on modern x86-64 platforms equipped with AES-NI and with vector AES-NI instructions. We explore several implementation methods and report a speed record for Rijndael256 at 0.27 cycles per byte. Keywords
Rijndael · Rijndael256 · AES · x86-64 · Software optimization · Wide block ciphers · AES new instructions · Vector instructions · Encryption throughput · Encryption
18.1
Introduction
Advanced Encryption Standard (AES) is the most ubiquitous symmetric cipher, and is used in many applications and scenarios. A prominent indication for its significance is the fast growing volume of AES-encrypted online data, which N. Drucker IBM Research, Haifa, Israel S. Gueron () University of Haifa, Haifa, Israel Amazon Web Services Inc., Seattle, WA, USA
18
is strongly supported by the industry (e.g., Intel’s AES-NI processor instructions [1, 2]). Recently, Intel added [3] new vectorized capabilities to the existing AES-NI instructions, namely VAESENC, VAESENCLAST, VAESDEC, and VAESDECLAST (VAES* for short). These instructions are intended to further speed up the performance of AES software, to a new theoretical throughput of 0.16 C/B. Software developers can use these instructions to speed up AES modes such as AES-Counter (CTR) and AES-CBC, as well as more elaborated schemes such as AES-GCM and AES-GCM-SIV [4] (a nonce misuse resistant AEAD). By definition, the block size of AES is 128 bits. Any 128-bit block cipher is a keyed permutation of {0, 1}128 . Therefore, regardless of the key length, ciphertext blocks can be distinguished (with high probability) from random strings, at the “birthday bound” of 264 blocks (in reality, way below that bound, when considering some security margins). Technically, the PRP-PRF distinguishing advantage is 1 2 128 q /2 when viewing q encryption samples. This limits 2 the number of encrypted blocks allowed with a given key. If we wish to keep the distinguishing advantage upper bounded by, say, 2−32 , the constraint 12 q2 /2128 ≤ 2−32 translates to re-keying after ∼248 blocks. While 248 blocks (=252 bytes) represent a huge amount of data, it is conceivable that current and near-future applications, especially at the cloud-scale, could benefit from lifting the required key rotation after 252 bytes. This motivates the exploration of a standardized 256-bit block cipher as a useful cryptographic primitive. Indeed, with 256 bits block size (permutation of {0, 1}256 ) the distinguishing advantage of 12 q2 /2256 poses no real constraints on the allowable number of processed blocks for any imaginable real scenario, even when the adversary advantage must remain below 2−32 . We point out that AES is based on the Rijndael proposal [5] that also defines a block cipher with 256 bits block size (and a 256-bits key), namely Rijndael256. However, only the
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_18
147
148
N. Drucker and S. Gueron
128-bit block size was standardized by the National Institute for Standards and Technology (NIST). Thus, Rijndael256 is a straightforward candidate to consider as a standardized 256-bit block cipher, provided that security analysis supports it as it supports AES (such analysis is still required, see for example [6]). Note that code that executes Rijndael256 can use the AES-NI, as shown in [2][Figure 30. “Using the AES instructions to compute a 256-bit block size RINJDAEL round”]. In 2016, Gueron and Mouha [7] reported that the throughput of code that uses AES-NI and executes pipelined Rijndael256 (encryption) are 1.54 cycles per byte (C/B hereafter). By comparison, the throughput of AES on the same processor is 0.65 C/B. With this in mind, we focus here on software optimizations for reaching the best achievable Rijndael256 performance on modern processors. Our Contribution We demonstrate for the first time the use of the new vector AES-NI extension to speed up the Rijndael256 block cipher in Electronic CodeBook (ECB) and CTR modes. We then measure our implementation on three different CPUs with different micro-architectures and analyze the results. Our results show that on modern CPUs the throughput of Rijndael256 is 0.27 C/B, which is comparable to the performance of the standard 128-bit block size AES. Organization The paper is organized as follows. Section 18.2 provides some background notation. Section 18.3 details our implementations. We report our experiments setup in Sect. 18.5 and performance results in Sect. 18.6. We conclude with Sect. 18.7.
18.2
Background and Notation
18.2.1 AES AES is defined in [8]. We describe it briefly, focusing on its encryption procedure (decryption is described in [8]). AES encryption takes a plaintext block of 128 bits and a key of size 128/192/256 bits as input and produces a 128bit ciphertext. The 128/192/256 bits key is expanded into 10/12/14 round keys, respectively. The plaintext is whitened by XOR-ring with the (first 128 bits) of the key, and then goes through 39/47/55 back-to-back transformations that can be organized as 9/11/13 identical AES rounds and an additional final round. The jth AES round, j=1, …, 9/11/13, is the sequence of transformations X = SubBytes(ShiftRows(S)) AESRound = MixColumns(X) ⊕ RoundKey[j]
operating on the 128-bits state S, where RoundKey[j] is the jth round key. The last round is the sequence X = SubBytes(ShiftRows(S)) AESLastRound = X ⊕ RoundKey[j] where j = 10/12/14.
18.2.2 AES-NI In 2009, Intel introduced six new instructions (AES-NI) to support AES computations. These instructions were and are defined in Gueron’s [1, 2] with samples and coding analysis. AESENC, AESENCLAST support encryption; AESDEC and AESDECLAST are building-blocks suitable for decryption (using the Equivalent Inverse Cipher); AESIMC and AESKEYGENASSIST support the Key Expansion. Since their introduction, the AES-NI extension has been added to most commodity processors (including AMD and ARM). Their use for fast AES encryption as well as other usages as well has become ubiquitous.
18.2.3 Vector AES-NI Recently, Intel has added a vectorized version of the AESNI (VAES*), as illustrated in Algorithm 1 (see [9] for details). These instructions can perform one round of AES encryption/decryption on KL = 1/2/4 128-bit operands (two qwords), having both register-memory and registerregister variant (we use only the latter here). The inputs are two source operands, which are 128/256/512-bit registers (named xmm, ymm, zmm, respectively), that (presumably) represent the round key and the state (plaintext/ciphertext). The special case KL = 1 using xmm registers degenerates to the current version of AES*. The vectorized instructions allow, theoretically, to quadruple the throughput of AES computations (compared to the xmm implementation), although the speed up my be lower on real processors, due to the execution-ports collisions when carrying out some needed shuffles and XOR operations.
18.2.4 Rijndael256 Rijndael256 [5] is a block cipher that receives as inputs a 256bit plaintext block and a 256-bit key. It treats the plaintext as a 4 × 8 matrix, where the 16 plaintext bytes p0 , . . . , p15 are mapped to the matrix elements a0,0 , a0,1 , . . . , a1,0 , . . . , a4,4 (see Fig. 18.3 for the matrix illustration).
18 Software Optimization of Rijndael256 for Modern x86-64 Platforms
Algorithm 1 VAES* instructions [3, 9] 1: 2: 3: 4: 5: 6: 7: 8: 9:
Inputs: SRC1, SRC2 (wide registers) Outputs: DST (a wide register) procedure VAES*(SRC1, SRC2) for i := 0 to KL − 1 do j = 128i RoundKey[127 : 0] = SRC2[j + 127 : j] T[127 : 0] = (Inv)ShiftRows(SRC1[j + 127 : j]) T[127 : 0] = (Inv)SubBytes(T[127 : 0]) T[127 : 0] = (Inv)MixColumns(T[127 : 0]) Only on VAESENC/VAESDEC. DST[j + 127 : j] = T[127 : 0] ⊕ RoundKey[127 : 0] return DST
The algorithm starts by expanding the key in a process called key scheduling to generate an array of 14 256-bit round keys ks. Figure 18.1 provides our key schedule implementation in C-style pseudocode. In a typical (software) application, the key is expanded once and then applied to some input stream of multiple blocks. Therefore, its effect on the overall process is relatively small. The Rijndael256 encryption and decryption flows are similar, and for brevity we only discuss the Rijndael256 encryption (see Algorithm 2). The Rijndael256 rounds are similar to the AES rounds and also use the ShiftRows, SubBytes, and MixColumns transformations. The difference between the AES and the Rijndael256 round lies in the bytes shuffling functionality, ShiftRows, where the two last rows of a given state are shifted by 3 and 4 columns, instead of by 2 and 3 columns as in AES. Details are provided, below, in the implementation Sect. 18.3.
149
The encryption round implementation of [1, 2] is based on AESENC and the Intel® AVX® architecture that includes 16 128-bit registers (xmm). This means that the Rijndael256 state is split across two xmm registers as described in Fig. 18.3 Panel (a). Applying two AESENC on these registers and in particular the ShiftRows transformation leads to the permutation in Panel (c). By contrast, the Rijndael256 ShiftRows transformation performs the permutation described in panel (d). To this end, the implementation of [2] performs a pre-permutation on the registers (panel (b)) before invoking the AESENC transformation. This is achieved through using the VPBLENDVB and PSHUFB instructions as in Fig. 18.2. Here, registers xmm1 and xmm2 hold the two halves of the Rijndael256 state (originally the input plaintext to encrypt), xmm6 and xmm7 hold the left half and right half of Rijndael256 round key, respectively. The updated (post-round) state is written into registers xmm1 and xmm2 and registers xmm8 and xmm5 hold the shuffling mask (0x03020d0c0f0e0908b0a050407060100) and the blend masks used for the pre-permutation (Fig. 18.3). On processors that have the AVX512 extension, we can store two 32-byte states in one wide 512-bit register (zmm). Subsequently, we can use the VPERMB instruction to replace the two serial operations VPBLENDVB and PSHUFB with one instruction and by that reduce the register dependency. For that, we use a mask with the following eight quad-words a0 , . . . , a7 . a0 a1 a2 a3 a4 a5 a6 a7
Algorithm 2 Rijndael256 block cipher [5] 1: 2: 3: 4: 5:
Inputs: state A 256-bit plaintext block, ks the key schedule Outputs: A 256-bit ciphertext block procedure Rijndael256(in, ks) state = in ⊕ ks[O] for i = 1; i < N ; i++ do state = Round(state, ks[i]) return LastRound(state, ks[14])
6: procedure Round(state, key) 7: state = Shif tRows(SubBytes(State)) 8: return M ixColumns(state) ⊕ key 9: procedure LastRound(state, key) 10: state = Shif tRows(SubBytes(State)) 11: return state ⊕ key
18.3
Our Implementations
Our implementation extends Gueron’s implementation of [2] and [1] that showed how to leverage AES-NI for Rijndael256. We describe our Rijndael256 key expansion [5][Section 3.6] implementation in Fig. 18.1
18.4
= 0x1b1a050417161100 = 0x13120d0c1f0e0908 = 0x0b0a151407060110 = 0x03021d1c0f1e1918 = 0x3b3a252437363120 = 0x33322d2c3f2e2928 = 0x2b2a353427262130 = 0x23223d3c2f3e3938
Rijndael256: The Best Possible Achievable Throughput
The AES algorithm (with a 128-bit key) consists of 10 AES rounds. Thus, the best AES throughput on a processor with AESENC on a single port, and with throughput 1, is ˜ 10/16 = 0.625 C/B. By contrast, Rijndael256 consists of a 32-byte state and 14 rounds, where every round involves two AESENC operations (every instruction only operates on half of the state), and some extra shuffling. Consequently, the best possible throughout is at best 2˜ ∗ 14/32 = 0.875. Better throughput is achievable when the platforms have AESENC on two execution ports. Note that the latency of the instruction only determines the number of independent blocks that we need to process in order to fill the pipeline.
150
N. Drucker and S. Gueron
Fig. 18.1 Rijndael256 key schedule in “C-style pseudocode”. Here, SLL and SHUF32 are short for _mm_slli_si128 and _mm_shuffle_epi32 C intrinsics, respectively
Fig. 18.2 Assembly code snippet from [2] that demonstrates the use of AES-NI for computing a Rijndael256 round
1 round1(in1, in2, idx, mask_ff) 2 __m128i t1 = AESKEYGEN(k, idx) 3 4 t1 = mask_ff ? SHUF32(t1, 0xff) : 5 SHUF32(t1, 0xaa) 6 7 __m128i t2 = SLL(in1, 0x4) 8 in1 ˆ= t2 9 t2 = SLL(t2, 0x4) 10 in1 ˆ= t2 11 t2 = SLL(t2, 0x4) 12 in1 ˆ= t2 ˆ t1 13 return *in1 14 15 intrlv_round(out[2], in[2], idx) 16 out[0] = round1(in[0], in[1], idx, 1) 17 out[1] = round1(in[1], in[0], 0, 0) 18 19 rijndael256_key_expansion(outKS, inKey) 20 vals[14] = {0x01, 0x02, 0x04, 0x08, 0x10, 21 0x20, 0x40, 0x80, 0x1b, 0x36, 22 0x6c, 0xd8, 0xab, 0x4d} 23 24 __m128i t[2] = {inKey[15:0], inKey[31:16]} 25 26 outKS[0] = t[0] 27 outKS[1] = t[1] 28 29 for(size_t i = 0; i < 14 ; i++) { 30 intrlv_round(outKS[2*(i+1)], t, vals[i]) 31 }
1 2 3 4 5 6
vpblendvb vpblendvb pshufb pshufb aesenc aesenc
Table 18.1 shows the latency and throughput of AESENC on different x86-64 processors [10]. Remark 18.1 (Counter Incrementation in CTR Mode) The (NIST specification) standard CTR mode is expressed in BigEndian convention. Therefore, on Little-Endian architecture (including x86-64 and ARM), the 4 bytes of the counter need to be swapped, incremented, and swapped back. Gueron and Krasnov [11] showed an optimization that alleviates this overhead for CTR mode. Their optimization was contributed to OpenSSL and is part of the CTR implementation thereof. For simplicity, in our studies, we encoded the counter using Little-Endian convention, where counters are incremented by simply using the 32-bit Single Instruction Multiple Data (SIMD) addition.
18.5
Experimental Setup
We carried out the performance experiments on three different platforms:
%xmm2, %xmm1, %xmm8, %xmm8, %xmm6, %xmm7,
%xmm1, %xmm5, %xmm3 %xmm2, %xmm5, %xmm4 %xmm3 %xmm4 %xmm3 %xmm4
• ICL: Dell XPS 13 7390 2-in-1 with the 10th Intel® CoreTM Generation (Micro architecture Codename “Ice Lake”[ICL]) Intel® CoreTM i7-1065G7 CPU 1.30 GHz. This platform has 16 GB RAM, 48K L1d and 32K L1i cache, 512K L2 cache, and 8MiB L3 cache. We turned off the Intel® Turbo Boost Technology (to work with a fixed frequency). • SKL: Lenovo P1 Gen 3 (type 20TH, 20TJ ) Laptop (ThinkPad)—Type 20TJ with the 8th Intel® CoreTM Generation (Micro architecture Codename “SkyLake”[SKL]) Intel® CoreT M i9-10885H CPU 2.40 GHz. This platform has 64 GB RAM, 32K L1d and 32K L1i cache, 128K L2 cache, and 2MiB L3 cache. On this platform, we could not turn off the Intel® Turbo Boost Technology (through the BIOS). Instead, we turned off the Hyper Threading capability and the power management feature of the CPU. • IVB: Lenovo T430s Laptop (Thinkpad) with the 3rd Intel® CoreTM Generation (Micro architecture Codename “Ivy Bridge”[IVB]) Intel® CoreTM i7-3520M CPU 2.90 GHz. This platform has 8 GB RAM, 64K L1d and
18 Software Optimization of Rijndael256 for Modern x86-64 Platforms
151
Fig. 18.3 The Rijndael256 state split to two registers xmm0, xmm1. Panel (a) is the initial state at the beginning of every round. Panel (c) show the results of permuting the two registers by applying the AES ShiftRows operation SR128() on the two registers. Panel (d) is the
desired Rijndael256 state. Green and white cells indicate correct and incorrect permuted cells. Panel (b) shows the pre-permutation required to go from (a) to (d). Gray cells were cross lanes between the registers
Table 18.1 The latency and throughput of AESENC and AESENCLAST on different processors [10]. Note the progress through the successive processor generations
• enc_lN—code based on [2] that uses AES-NI intrinsics, where we use the GCC/Clang unroll loops attribute (i.e., _Pragma(”GCC unroll N“) and _Pragma(“clang loop unroll_count(N)”)) for GCC and Clang, respectively) with an unroll parameter N. • enc_xN—encryption function that processes N (independent) blocks in parallel. The code is based on [2], where we interleave independent operations in order to get perfectly full pipelining. This is done by duplicating every original-code-row N times. • vaes—Code that uses the VAES* instructions, and processes two blocks at a time. • vaesxN—Code that uses the VAES* instructions and processes 2N (independent) blocks in parallel. The code interleaves independent operations as above.
Architecture Ivy bridge [IVB] Haswell [HSW] Broadwell [BDW] Skylake [SKL] IceLake [ICL]
Latency 8 7 7 4 4
Throughput 1 1 1 1 0.5
64K L1i cache, 256K L2 cache, and 1MiB L3 cache. We turned off the Intel® Turbo Boost Technology.
18.5.1 Our Optimized Code We wrote our code in C, using x86-64 C intrinsics. The implementation uses the AVX512 and VECTOR-AES instructions. The code was compiled with both GCC (version 10) and Clang (version 12) compilers, in 64-bit mode, and with the “O3” optimization level (i.e., -O3 compilation flag), and run on a Linux OS (Ubuntu 20.04.2 LTS). For the study, we compared several Rijndael256 implementations, encrypting data in ECB and in CTR modes. The explored code options are: • enc—code based on [2] that uses AES-NI intrinsics.
18.5.2 Measurements Methodology The performance measurements are reported here in cycles per byte (C/B), and the executables’ performance is profiled by using the RDTSCP instruction (per single core). Lower cycles count indicates faster execution (i.e., better performance). We used the following measurement methodology in our experiments. Every measured function was isolated, run 100 times (warm-up), followed by 1000 iterations that were measured and averaged. To minimize the effect of background tasks running on the system, we repeated every experiment 10 times and recorded the minimal result.
152
18.6
N. Drucker and S. Gueron
Performance Results
The performance results are shown in Fig. 18.4 and Tables 18.2, 18.3. on ICL, our new vectorized implementation is 5.8× faster than the non-vectorized implementation. In addition, we see that the loop unrolling attribute does not affect the final throughput. This indicates that the tested compilers cannot (automatically) fully pipeline the code. The SKL
processor has AESENC on only a single execution port. Here, we report throughput of 0.7 C/B. This is even better than our 0.875 C/B estimation, and our explanation for that is that we could not turn off the Intel Turbo feature on that platform (and hence received some extra performance boost). These results give us the performance for systems that are configured in this way.
Fig. 18.4 The performance (in C/B; lower is better) of Rijndael256 in ECB and CTR modes, measured on Ice Lake processor. The results show different code implementations (see the text for details), compiled under GCC-10 and Clang-12. Different compilers achieve different performance results on the same code. The best performance of 0.27 C/B is achieved by using vector AES, pipelining 8 blocks, and compiling with GCC-10
Table 18.2 The performance (in C/B; lower is better) of Rijndael256 in ECB and CTR modes, measured on ICL processor. The table provides detailed information from Fig. 18.4. The boldface numbers show the best Method enc enc_l2 enc_l4 enc_l8 enc_x2 enc_x4 enc_x8 vaes vaesx2 vaesx4 vaesx8
ECB mode gcc-10 2.28 2.26 2.24 2.24 2.12 1.57 1.73 0.42 0.38 0.31 0.27
clang-12 3.22 3.2 3.21 3.21 2.79 2.04 2.07 0.4 0.35 0.29 0.27
performance per column. The results show that aggressive parallelization with CTR mode brings the performance close to the ECB mode that is the fastest possible (though only “theoretical”) mode
CTR mode gcc-10 2.42 2.42 2.41 2.41 2.2 1.6 1.84 0.4 0.37 0.31 0.29
gcc-clang speedup 1.41 1.42 1.43 1.43 1.32 1.30 1.20 0.95 0.92 0.94 1.00
clang-12 3.29 3.28 3.28 3.3 2.82 2.03 2.06 0.4 1.2 0.8 0.53
gcc-clang speedup 1.36 1.36 1.36 1.37 1.28 1.27 1.12 1.00 3.24 2.58 1.83
CTR-vs-ECB 1.06 1.07 1.08 1.08 1.04 1.02 1.06 0.95 0.97 1.00 1.07
Table 18.3 The performance (in C/B; lower is better) of Rijndael256 in ECB and CTR modes, measured on SKL (left) and IVB (right) processors. The boldface numbers show the best performance per column Method enc enc_l2 enc_l4 enc_l8 enc_x2 enc_x4 enc_x8
ECB mode gcc-10 1.1 1.09 1.06 1.05 0.87 0.7 0.91
clang-12 1.21 1.2 1.21 1.21 0.92 0.86 0.91
CTR mode gcc-10 1.1 1.12 1.09 1.12 0.85 0.78 0.92
clang-12 1.23 1.24 1.26 1.23 0.95 0.87 0.92
Method enc enc_l2 enc_l4 enc_l8 enc_x2 enc_x4 enc_x8
ECB mode gcc-10 4.41 4.35 4.34 4.31 3.36 2.62 2.39
clang-12 4.12 4.08 4.07 4.06 2.96 2.46 2.23
CTR mode gcc-10 4.38 4.27 4.24 4.25 3.43 2.67 2.59
clang-12 4.18 4.21 4.13 4.09 2.97 2.33 2.23
18 Software Optimization of Rijndael256 for Modern x86-64 Platforms
18.7
Conclusion
As anticipated, we see that code optimizations need to take the specific processor characteristics into account. In our case, the performance is dominated by the AESENC instruction latency, its throughput, and the level of parallelization offered by the architecture (AVX, AVX512). Table 18.1 shows the information per the different processors. Other factors that affect the performance are the counter increments, shuffles, blends, XOR operations and memory access. Our code is optimized (per architecture) for these considerations as well. It is interesting to see that the compilers (GCC & Clang) behave differently when requested to automatically pipeline vectorized code (for leveraging VAESENC). In our studies, we found that GCC outperforms Clang in most cases. In the end, we were able to write optimized code that executes Rijndael256 at the rate of 0.27 C/B (on ICL), 0.27 C/B (on SKL), and 2.23 C/B (on IVB). The speed record of 0.27 C/B on the latest processor generation (ICL) indicates that the 256-bit block cipher can be implemented in software in an extremely efficient way. It is interesting to compare the performance for the 256-bit block cipher Rijndael256 to that of the standard AES (with 128-bit blocks). We take one case to estimate the cost of doubling the block size. We focus on parallelizable modes, CTR in particular, nonvectorized implementation pipelining 8 blocks: AES-CTR with a 256-bit key performs at 0.58 C/B, respectively on ICL, and we compare this to AES-Rijndael256 (with a 256bit key) that performs at 1.84 C/B on the same processor, showing a ∼3.3× factor. The wide SIMD architectures alleviate the relative overheads of the shuffling and blending and reduce the relative performance impact of the doubled block size.
153
We made our code available as an open-source sample under the Apache-2.0 license and uploaded it to the GitHub repository https://github.com/shay-gueron/Rijndael256. Acknowledgments This research was partly supported by: NSF-BSF Grant 2018640; The Israel Science Foundation (grant No. 3380/19); The Center for Cyber Law and Policy at the University of Haifa, in conjunction with the Israel National Cyber Bureau in the Prime Minister’s Office.
References 1. S. Gueron, Intel® Advanced Encryption Standard (AES) New Instructions Set Rev. 3.01. Intel Software Network (2010) 2. S. Gueron, Intel’s new aes instructions for enhanced performance and security, in ed. by O. Dunkelman, Fast Software Encryption (Springer, Berlin, 2009), pp. 51–66 3. Intel. Intel® 64 and IA-32 architectures software developer’s manual, June (2021) 4. S. Gueron, A. Langley, Y. Lindell, AES-GCM-SIV: Nonce MisuseResistant Authenticated Encryption. RFC 8452 (2019) 5. J. Daemen, V. Rijmen, The Design of Rijndael, AES – The Advanced Encryption Standard (2002) 6. Y. Liu, Y. Shi, D. Gu, B. Dai, F. Zhao, W. Li, Z. Liu, Z. Zeng, Improved impossible differential cryptanalysis of large-block Rijndael. Sci. China Inf. Sci. 62(3), 32101 (2018) 7. S. Gueron, N. Mouha, Simpira v2: A family of efficient permutations using the AES round function, in ASIACRYPT (1). Lecture Notes in Computer Science, vol. 10031 (Springer, Berlin, 2016), pp. 95–125 8. NIST. FIPS pub 197: Advanced encryption standard (AES) (2001) 9. N. Drucker, S. Gueron, V. Krasnov, Making AES great again: The forthcoming vectorized AES instruction, in ed. by S. Latifi, 16th International Conference on Information Technology-New Generations (ITNG 2019) (Springer International Publishing, Cham, 2019), pp. 37–41 10. Intel. Intel® intrinsics guide, Oct (2021) 11. S. Gueron, V. Krasnov, The fragility of AES-GCM authentication algorithm, in ITNG (IEEE Computer Society, Washington, 2014), pp. 333–337
Cybersecurity Ethics Education: A Curriculum Proposal
19
Ping Wang
Abstract
Cybersecurity ethics has emerged as a new and increasingly significant area for research and education. New and complex ethical dilemmas and conflicts in values and judgments arise as new cybersecurity technologies and policies are constantly explored and implemented to defend our cyberspace for a safe and secure living and work environment. As various cyber threats, attacks, and risks pose increasing challenges to the diverse and interconnected world we live in, there is an increasing demand for quality cybersecurity education to prepare and produce qualified and ethically competent professionals to address the cybersecurity challenges. The national Centers of Academic Excellence in Cyber Defense Education (CAECDE) designation by the U.S. National Security Agency and Department of Homeland Security (NSA/DHS) is a high-quality program that promotes excellence in cybersecurity education for developing qualified cyber talent. Strong curriculum and courses supported by regular mentoring are essential to successful preparation of cybersecurity professionals. This research paper proposes a new credit course in cybersecurity ethics supported by an adopted comprehensive model of mentoring for an undergraduate cybersecurity education program at a CAE-CDE designated university in the United States. The curriculum proposal will present the rationale, course description, mappings of learning outcomes and topics to the CAECDE knowledge unit, suggested methods of assessment, and mentoring activities. The goal of this research is to contribute a new course design with ethical mentoring to P. Wang () Robert Morris University, Pittsburgh, PA, USA e-mail: [email protected]
enrich and enhance national and international cybersecurity curriculum and education.
Keywords
Cybersecurity · Ethics · Curriculum · CAE-CDE · Knowledge unit · Mentoring
19.1
Introduction
We work and live in a digitally connected world with increasing and more complex cyber threats, attacks, and risks. Accordingly, there has been a substantial shortage of and increasing workforce demand for well-educated and qualified cybersecurity professionals to defend our cyber space and critical assets [1, 2]. A recent cybersecurity workforce survey study shows that the shortage of cybersecurity professionals is nearly three million across the world and about half a million in North America and the majority of organizations reported concerns of risks of cybersecurity attacks due to insufficient cybersecurity staff [3]. In addition, the cybersecurity domain is faced with increasingly complex ethical dilemmas and conflicts of values and judgement, such as security versus privacy, security versus accountability, security versus social equality and fairness, security restrictions versus civil liberties, etc. [4]. There has been a lack of awareness of and inadequate guidance on cybersecurity ethics in both research and practitioner communities [5]. Higher education providers are expected to prepare and produce qualified cybersecurity professionals to meet the workforce demand and ethical challenges in the cybersecurity field. However, a realistic challenge for cybersecurity
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_19
155
156
education is to ensure that the graduates are qualified for the expectations of cybersecurity jobs in terms of knowledge, skills, and abilities (KSAs). The U.S. National Initiative for Cybersecurity Education (NICE) published the NICE Cybersecurity Workforce Framework (NCWF) and recent revisions to provide a standard taxonomy and lexicon for specific cybersecurity work categories, job roles and tasks for the cybersecurity industry and corresponding expectations of knowledge, skills, and abilities (KSAs). NCWF serves as a guidance for employers, education and training providers, and public and private industry sectors to define cybersecurity work categories, job roles, and their required professional skills [6–8]. The NICE framework includes the knowledge of ethics related to cybersecurity and privacy as an essential professional qualification for all cybersecurity work roles [6]. A credible and rigorous quality assurance mechanism is needed to evaluate and maintain the program quality of cybersecurity education and training providers. The national Centers of Academic Excellence in Cyber Defense Education (CAE-CDE) designation program in the United States jointly sponsored by the National Security Agency (NSA) and Department of Homeland Security (DHS) has been the most comprehensive and reputable national standard for evaluating, certifying, and maintaining the quality of cybersecurity education programs. The CAE-CDE program has established specific and measurable criteria for program evaluation and assessment of cybersecurity knowledge units. The knowledge units (KUs), including the Cybersecurity Ethics KU, are a key component of the CAE-CDE designation program quality standard that applies to all cybersecurity degree programs for designation [9]. The CAE-CDE KUs and associated learning outcomes and topics are valuable reference guide for developing strong cybersecurity curriculum and courses. Effective implementation and delivery of a strong cybersecurity program will benefit from supporting mentoring activities. Prior research has indicated that mentoring is an important factor for student academic and professional success in cybersecurity education [1, 10, 11]. A comprehensive mentoring model that includes academic mentoring, career guidance, extra-curricular mentoring, and ethical guidance is found to have contributed to student academic and professional success in cybersecurity student development [1, 12]. The following sections of this article will review relevant theoretical background on cybersecurity ethics education and mentoring, present a cybersecurity ethics curriculum proposal using the case study approach, describe and discuss the cybersecurity ethics course design and curriculum mappings, and provide conclusions and suggest future directions for follow-up research in this area.
P. Wang
19.2
Background
Ethics is a set of values for making judgements and decisions within a certain group, such as a cultural, social, or professional group, on acceptable lifestyle or business conduct. Ethical values may agree or conflict with other values and developments. Cybersecurity aims to protect and secure certain valued properties and assets in the digital space. However, cybersecurity technologies and policies may cause conflicts with certain ethical values in both private and public sectors. Ethical issues are wide-ranging in the private business domain. The most frequently discussed ethical issues in business cybersecurity research literature include privacy, data protection, trust, control, accessibility, confidentiality, ethical codes of conduct, data integrity, informed consent, transparency, availability, and accountability [13]. However, additional research literature is needed to provide thorough analysis of businesses’ ethical responsibilities to stakeholders in the contexts of cybersecurity incidents, such as a ransomware attack or a business data breach [13]. Ethical dilemmas and conflicts of values may also occur in the public domain cybersecurity, such as in securing critical infrastructures as part of national security. A potential conflict of values is between the privileged access or backdoors of government agencies to ICT services may jeopardize the essential values of privacy and freedom for a democratic society; For example, the use of artificial intelligence (AI) as a defensive and preventive cybersecurity technology in law enforcement, national security, and counter-terrorism may lead to faster detection, identification, and response to threats, but it may expose civilians to the threat of authoritarian government abuse of the technology in the cyber domain [14]. Cybersecurity professional organizations have adopted code of ethics to emphasize the importance of ethics in cybersecurity professional practice. For example, the International Information System Security Certification Consortium, or (ISC)2 , has adopted the following four mandatory Code of Ethics Canons as high-level guidance for (ISC)2 certified professionals: • Protect society, the common good, necessary public trust and confidence, and the infrastructure. • Act honorably, honestly, justly, responsibly, and legally. • Provide diligent and competent service to principals. • Advance and protect the profession [15]. ISACA (Information Systems Audit and Control Association) has adopted comparable code of professional ethics for its members and certification holders [16].
19 Cybersecurity Ethics Education: A Curriculum Proposal
Due to the ethical challenges for computing and cybersecurity technology, research in computing and cybersecurity education has also recognized the growing importance of introducing and integrating ethics education in cybersecurity curricula to train students to be aware of the ethical dilemmas and capable of making ethical judgment and decisions [17– 19]. Ethical commitment to one’s organization and national security is one of the six key personal traits needed for professional success in the future cybersecurity workforce [20]. Cybersecurity education is multidisciplinary involving computing science and technology, management, communication, critical thinking, problem-solving, and analytical and decision skills [1]. Ethics component should be included and integrated into a cybersecurity curriculum for ethical judgement and decision making. The United States Naval Academy has adopted a comprehensive and interdisciplinary approach to cybersecurity education that successfully blends technical courses such as programming and networking with non-technical courses such as policy, economics, law, and ethics [21]. To complement ethics course work in the cybersecurity curriculum, ethical mentoring and guidance should be used to support the development of student ethical competency. An effective mentoring relationship between faculty mentors and student mentees will guide the student mentees toward success. Student success is defined as both academic success in completing the academic program and professional success in gaining readiness for the cybersecurity profession. The comprehensive mentoring model in Fig. 19.1 below, which includes ethical guidance, serves to help students to achieve both academic success and professional success. The four mentoring components are interrelated and work together to contribute to students’ overall success in cybersecurity education and professional preparation. Mentoring in ethics may involve the following topics and activities: • Ethical professional behavior • Seminar sessions on the importance of authorization in penetration testing • Data security and privacy laws • Compliance and reporting requirements • Non-disclosure agreement • Security clearance requirements • Conflict of interest disclosures • Copyright protection and plagiarism • Ethics in cybersecurity research Mentors should also guide student mentees and make them aware of important code of ethics for major cybersecurity professional groups and societies, such as the mandatory Code of Ethics for professionals certified by (ISC)2 and ISACA [15, 16].
157
Career Guidance
Academic Mentoring
Student Success
Extracurricular Mentoring
Ethical Guidance
Fig. 19.1 Comprehensive mentoring model for cybersecurity education [1]
19.3
Curriculum Proposal
This research presents a sample cybersecurity curriculum proposal using the case study methodology. The case study is based on the curriculum of a Bachelor of Science (BS) degree program in cybersecurity at a CAE-CDE designated university in the northeast of the United States. The 123-credit degree program is also accredited by ABET-CAC and already has well-rounded technical and non-technical courses for a 4year sequence, but the curriculum does not have a course in cybersecurity ethics yet. This research is to propose a 3-credit senior level seminar course in cybersecurity ethics.
19.3.1 Rationale The course in cybersecurity ethics is necessary for graduates of the program to be aware of the complex ethical dilemmas and be able to make sound ethical judgment and decisions in the cybersecurity profession. This is in support of the ABET accredited cybersecurity program learning outcome in ethics: Recognize professional responsibilities and make informed judgments in computing practice based on legal and ethical principles.
158
19.3.2 Course Description This is a senior seminar that addresses important ethical issues, challenges, frameworks, and best practices in cybersecurity. The course work includes substantial discussions of readings and case studies to develop strong competency in ethical standards and judgment for cybersecurity decisions and behavior for service in public and private sectors.
19.3.3 Course Learning Outcomes The course learning outcomes are mapped to the outcomes for the Cybersecurity Ethics (CSE) knowledge unit for the CAE-CDE designation, which expects students to be able to: (1) Explain how ethical foundations are applied to situations arising from the interconnected world; (2) Examine diverse ethical dilemmas; and (3) Describe the role of cybersecurity in supporting and encouraging ethics, as well as where cybersecurity practices can cause ethical conflicts [9].
19.3.4 Course Topics The course topics are mapped to the suggested topics in the CSE KU for CAE-CDE programs, which include the following [9]: • Ethical codes and frameworks • Ethics and cyberspace • Ethical issues, such as property, rights of others, availability, respect and principles of community, resource use, allocation, and abuse, and censorship • Ethics-based decision tools • Cybersecurity and social responsibility
19.3.5 Assessment Methods Assessment of student learning is essential to measure students’ achievement in knowledge, skills and abilities and necessary for continuous improvement of curriculum and course design. Assessment activities for the cybersecurity ethics course may utilize the following activities and assignments for evaluation: • • • • • •
Case studies Class discussions Individual essay assignments Individual research assignment Team project and presentation Midterm and final exam
P. Wang
These activities will measure student knowledge and skills and abilities to analyze, evaluate and make judgement on the cybersecurity ethics topics covered in the course.
19.4
Ethical Mentoring
To support student learning in the proposed course in cybersecurity ethics, faculty mentoring of students is necessary, which is based on the Ethical Guidance component in the comprehensive mentoring model [1]. In addition to the formal coursework in Cybersecurity Ethics done in the cybersecurity program, students will receive continuous ethical training and guidance during their junior, senior, and graduate years from faculty mentors and government and business partnerships and volunteers to develop a strong ethical character and behavior for public or corporate service. The group and individual ethical guidance will include presentations and discussions led by faculty and government and corporate cybersecurity professionals on various ethical issues such as cybersecurity professional codes of conduct, privacy protection, conflict of interest, security clearances, and specific ethics policies and expectations for federal, state and local government services.
19.5
Conclusions
There is increasing demand for qualified cybersecurity workforce to defense our cyberspace from various threats and attacks. Ethical issues are increasingly significant in the cybersecurity field with complex dilemmas and challenges for cybersecurity professionals. Cybersecurity education programs are expected prepare and produce high-quality and ethical cybersecurity professionals. This research presents a case-based curriculum proposal including a cybersecurity ethics course supported by ethical mentoring for student success. The curriculum and course design is mapped to the current Cybersecurity Ethics Knowledge Unit for the U.S. national CAE-CDE designation and ABET-CAC cybersecurity program accreditation. The supplemental ethical mentoring is based on the comprehensive mentoring model from recent research in cybersecurity education. This study is preliminary on the topic of curriculum design for cybersecurity ethics education. The proposed course design is limited to the outline of the course description, learning outcomes, topics, and assessment. More data will be available after implementations of the proposed curriculum design. Future research directions on this topic may include the impact of the coursework on student knowledge, skills and abilities on cybersecurity ethics and the effectiveness of the instructional and mentoring activities.
19 Cybersecurity Ethics Education: A Curriculum Proposal Acknowledgement This research was supported by a grant award from the United States Department of Defense Cyber Scholarship Program (Grant No. H98230-20-1-0352).
References 1. P. Wang, R. Sbeit, A comprehensive mentoring model for cybersecurity education, in Advances in Intelligent Systems and Computing, ed. by S. Latifi, (Springer Nature Switzerland AG, 2020), pp. 17–23 2. U.S. Department of Labor BLS (Bureau of Labor Statistics), (2021). Retrieved from https://www.bls.gov/ooh/computer-andinformation-technology/information-security-analysts.htm 3. (ISC)2, Cybersecurity Professionals Focus on Developing New Skills as Workforce Gap Widens: (ISC)2 Cybersecurity Workforce Study 2018, (2018). Retrieved from https://www.isc2.org/research 4. M. Christen, B. Godijn, M. Loi, The ethics of cybersecurity, in The International Library of Ethics, Law and Technology, vol 21, (SpringerOpen, 2020) 5. K. Macnish, J. van der Ham, Ethics in cybersecurity research and practice. Technol. Soc. 63(101382), 1–10 (2020) 6. NICE (National Initiative for Cybersecurity Education), NICE Cybersecurity Workforce Framework (SP800-181), (2017). Retrieved from https://nvlpubs.nist.gov/nistpubs/SpecialPublications/ NIST.SP.800-181.pdf 7. NICE (National Initiative for Cybersecurity Education), NIST (National Institute of Standards and Technology), NICE Cybersecurity Workforce Framework (SP800-181 Revision 1), (2020). Retrieved from https://doi.org/https://doi.org/10.6028/NIST.SP.800-181r1 8. P. Wang, M. Dawson, K.L. Williams, Improving cyber defense education through national standard alignment: Case studies. Int. J. Hyperconnect. Internet Things 2(1), 12–28 (2018) 9. NIETP (National IA Education & Training Programs), 2020 CAE Cyber Defense (CAE-CD) Knowledge Units, (2021). Retrieved from https://public.cyber.mil/ncae-c/documents-library/ 10. K. Dean, What everybody ought to know about mentoring in InfoSec. AT&T Cybersecurity, (2019). Retrieved from https:// www.alienvault.com/blogs/security-essentials/what-everybodyought-to-know-about-mentoring-in-infosec
159 11. E. Dallaway, How to mentor an information security professional. Infosecurity Magazine, (2016). Retrieved from https:// www.infosecurity-magazine.com/blogs/how-to-mentor-aninfosec-pro 12. P. Wang, Cybersecurity student talent recruitment and development: A case study. Issues Inf. Syst. 22(2), 210–222 (2021) 13. G. Morgan, B. Gordijn, A care-based stakeholder approach to ethics of cybersecurity in business, in The Ethics of Cybersecurity. The International Library of Ethics, Law and Technology, vol 21, ed. by M. Christen, B. Godijn, M. Loi, (SpringerOpen, 2020), pp. 119–138 14. E. Vigano, M. Loi, E. Yaghmaei, Cybersecurity of critical infrastructure, in The Ethics of Cybersecurity. The International Library of Ethics, Law and Technology, vol 21, ed. by M. Christen, B. Godijn, M. Loi, (SpringerOpen, 2020), pp. 157–178 15. (ISC)2 , (ISC)2 Code of Ethics, (2021). Retrieved from https:// www.isc2.org/Ethics 16. ISACA, Code of Professional Ethics, (2021). Retrieved from https:/ /www.isaca.org/credentialing/code-of-professional-ethics 17. L. Adaryukova, O. Bychkov, M. Kateryna, A. Skyrda, The introduction of ethics into cybersecurity curricula, in Proceedings of the 16th International Conference on ICT in Education, Research and Industrial Applications, Kharkiv, Ukraine, October 6–10, 2020. 25–36, (2020) 18. J. Blanken-Webb, I. Palmer, S.-E. Deshaies, N.C. Burbules, R.H. Campbell, M. Bashir, A case study-based cybersecurity ethics curriculum, in Workshop on Advances in Security Education (ASE 18), (USENIX) Association, Baltimore, MD, (2018) 19. Grosz et al., Embedded ethics: Integrating ethics across CS education. Commun. ACM 62(8), 54–61 (2019) 20. J. Dawson, R. Thomson, The future of cybersecurity workforce: Going beyond technical skills for cyber performance. Front. Psychol. 9(744), 1–12 (2018) 21. T. Emmersen, J.M. Hatfield, J. Kosseff, S.R. Orr, The USNA’s interdisciplinary approach to cybersecurity education, in Computer, (IEEE Computer Society, 2019 March), pp. 48–57
Performance Evaluation of Online Website Safeguarding Tools Against Phishing Attacks; a Comparative Assessment
20
Rama Al-Share, Fatima Abu-Akleek, Ahmed S. Shatnawi, and Eyad Taqieddin
Abstract
Despite the security policies that organizations follow to defend against cyber crimes, phishing attacks are still among the most popular ways the criminals use to steal user’s credentials. Spear phishing, fake websites, fraudulent emails, smishing, and vishing all fall under the umbrella of phishing attacks. Recent procedures followed by many organizations tend to develop anti-phishing tools that identify fraudulent emails and websites, which are embedded either implicitly within the web browsers and email applications or explicitly as an online service. In this research, we have evaluated the effectiveness of six online checking tools in detecting potentially malicious websites. Six URL scanning engines from the best wellknown engines in the VirusTotal website were tested on a list of legitimate and malicious URLs, which were collected from well-known anti-phishing frameworks, including PhishTank. In order to find the most efficient antiphishing tool, the detection accuracy, precision, recall, and F1-Score were calculated for each engine. The results showed that among website checking tools, Sophos achieved a detection accuracy of 99.23% and a precision value of 100%. Keywords
Phishing attacks · Anti-phishing tools · Online checkers · Malicious website detection
R. Al-Share · F. Abu-Akleek · A. S. Shatnawi () · E. Taqieddin Jordan University of Science & Technology, Irbid, Jordan e-mail: [email protected]; [email protected]; [email protected]; [email protected]
20.1
Introduction
One of the most common risks that have been defined by the Open Web Application Security Project (OWASP) is the broken authentication and session management vulnerability, which comes in the second-order of the OWASP top 10 vulnerabilities since 2013. By leveraging this type of vulnerability, the attacker could steal credentials of legitimate users (i.e., usernames and passwords) or even hijack their web sessions to impersonate their identities and login to the system with valid credentials [1]. The organizations that use weak authentication and session management methods are vulnerable to this type of data breach. Over several years, phishing attacks were classified as the highest critical fraud attacks, which fall under the broken authentication security risks. The attacker usually tries to fool some users using social engineering techniques to steal their data, including login credentials or credit card information. From a personal perspective, stealing any personal information from users readily violates their privacy and may lead to impersonation attacks. Once an attacker obtains user credentials, he can impersonate their identity to launch different cybercrimes like making illegitimate financial transactions to steal money from the user’s bank accounts. From a business perspective, the reputation of organizations could be severely affected as a result of phishing attacks. For instance, when a well-known and trusted financial institution is subjected to a certain data breach involving credit card information for one million users, all at one time, customers will lose trust in such organization: hence, affecting its productivity and reputation in the market. The organizations will lose millions of dollars for multiple years in order to heal from this financial impact.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_20
161
162
R. Al-Share et al.
Sect. 20.5, where Sect. 20.6 provides the conclusion and future work commensurate with this research.\include
Vishing and smishing are also popular types of phishing attacks, in which an attacker uses voice calls and SMS messages, respectively, to fool the users. In smishing, the attacker tries to send professional SMS messages to the victims convincing them to click on the given malicious link in order to validate their identity. Usually, the attacker impersonates the identity of a trusted bank or organization to deceive victims. In vishing, on the other hand, the attacker aims at tricking the users through voice calls, persuading them that they are calling from a trusted organization, so they must provide their identity number, credit card information, or any other sensitive information. This information is used by the attacker, later, to steal money or compromise the user account. Due to the highly critical risks caused by phishing attacks, many countermeasures are taken by various organizations to defend against them, including the activation of the Two Factor Authentication (2FA) property, filtering emails and websites, and the use of anti-phishing tools. Usually, the antiphishing tools could be used either implicitly within web browsers and email applications or explicitly as an online service. In this research, we study the effectiveness of online checking tools in detecting phishing websites. The study is performed on six online detection tools and were tested against two types of website datasets to identify phishing websites. A comparison between the determined detection tools was made opposite the detection accuracy and other evaluation metrics. The rest of this paper is organized as follows: Sect. 20.2 describes the scenario of phishing attacks. Section 20.3 provides a literature review of existing phishing detection schemes. Section 20.4 presents the online checking tools that are presented in this research. The evaluation procedure, dataset description and performance results are revealed in
Attack Description
Generally, attackers start to launch phishing attacks by sending many fake emails to targeted or untargeted users. These phishing emails usually contain malicious URLs that lead the users to phishing websites designed specifically to steal user’s credentials and access personal accounts. The most important factor that makes the operation more successful is the human factor, which is used by the attackers to collect information about users and professionally write emails that sound appealing from a user’s point of view. This information could represent the user’s interests, names of relatives and family members, or even the work environment the user is associated with. The whole operation of launching phishing attacks falls under the umbrella of social engineering attacks [2]. Figure 20.1 describes the attack scenario and the following stages describe specifically the operation of phishing attacks:
20.2.1 Send Phishing Emails to the Users In the initial stage, the attackers prepare a list of phishing emails and send them to many users. Usually, at first look, the format of these emails looks interesting to the users, so they will be deceived to open the malicious URLs carried in the body of the message. At this stage, the attacker uses the information collected about the users and, thereon, applies social engineering techniques to write phishing emails.
Attacker
User Step1: Send phising emails to users
Step4: Redirect credentials to attacker
Fig. 20.1 The scenario of phishing attacks
20.2
p2:
Ste
ire Red
ct t
om
3: R
a
equ
us licio
e
re st c
web
den
site
tials
p Ste
Step4: Redirect credentials to attacker Phishing Website
20 Performance Evaluation of Online Website Safeguarding Tools Against Phishing Attacks; a Comparative Assessment
20.2.2 Redirect the Users to Malicious Websites When the users click on the malicious links provided within the phishing email’s content, they will be redirected to malicious websites created specifically to steal user’s credentials or other secret information. These websites are carefully designed so that the users would be inclined to believe that the website is authentic.
20.2.3 Request Credentials from Users As the legitimate website does, the malicious website requests from the users to enter their credentials, like usernames, passwords, and even credit card information. This stage is the most dangerous and undetectable action in phishing attacks, especially when the design of the website and the URL look the same as on the legitimate one. As a result, a user must be able to acquaint him/herself of the minute differences in the contents and the provided URL before releasing any secret credentials to the fake website.
20.2.4 Forward Credentials to the Attacker Since the malicious website is created by an attacker, all the given information will be transferred to the attacking device. The attacker will be able to access the user’s account and reveal any other secret information if he is successful in taking control of the username and password of the user account. Also, if he could obtain the credit card information, he will be able to transfer money from the user’s bank account to his own. The actions depend on the attacker’s specific aims and the activities he is associated with.
20.3
Literature Review
In this section, we present a literature review of existing detection solutions and recent survey papers which have addressed the area of phishing attack detection. \include Authors in [3] propose a summary and classification of phishing attacks. Further, they also address certain types of phishing attacks and provide a discussion of some techniques and countermeasures to defend against them. The authors classify phishing into two categories: social engineering such as fake websites and e-mail spoofing, and technical ploy such as cross site scripting, malware phishing and session hijacking. Therein, authors have relied on two classifications in order to classify the defense solutions, the email filtering schemes and the websites phishing schemes. There are some
163
effective features used to discover email spoofing including body-based, header-based, URL-based and sender-based features. Pursuant with that, the authors classify email filtering into A. Network level protection, which blocks a set of known malicious domains or IP addresses as examples of it. In order to mitigate the phishing websites authors also classify the solutions into A. Blacklist and whitelist where these lists include the malicious and legitimate IPS and URLs, as examples of this method Google Safe Browsing API, PhishNet: Predictive Blacklisting, and Automated Individual White-List. B. Heuristic Solutions which solve the problems based on previous results and experiences, examples of this method as SpoofGuard tool, Collaborative Intrusion Detection system (CIDS), PhishGuard: A Browser Plug-in, and CANTINA: A Content-Based Approach. C. Visual Similarities such as Visual Similarity Based Phishing Detection, and BaitAlarm. D. Miscellaneous Solutions that not belong to the previous categorize such as TrustBar and Dynamic Security Skin. An empirical study was conducted in [4] to determine the main properties that are most vulnerable in webpages, to improve the selection process the authors relayed on machine learning techniques, called bagging, using the Weka program; here, the study is conducted using data of phishing and botnet malwares. Authors depend on the random tree as a classifier since it gives the best result to a high degree of accuracy; the phishing dataset include phishy behavior, suspicious behavior and legitimate behavior. While phishy behavior which aims to attack the users and steal their sensitive information to do harm, suspicious behavior is an activity which could be harmful or phishy include some malicious codes and links, and a legitimate one which is the normal case and does not contain any harmful links or codes. The authors reported on eleven features to analyze the phishing websites malware behavior some of which are IP address and port number, URL and google index, web traffic and HTTP tokens, and links pointing to pages. The dataset was analyzed, and the results were sent to an Excel sheet to make charts and draw graphs of the results. Results show that google indices are more secure to browse where just 13% of them were classified as phishy behavior, in addition to achieving an accuracy of 89% for both phishing and botnet datasets on average. Authors in [5] propose a technique to detect phishing websites based on URL and domain name features where it can classify the websites into legitimate and fake ones. They dubbed their designed program by PhishChecker The authors set a group of rules to determine which of the websites were phished and which were legitimate; these rules were
164
A. if the IP address is included within the URL then it implies a phishing webpage. B. if the length of the URL is greater than 54 characters then a phishing website is identified. C. if the URL contains dash mark (−) then a phishing website is again identified. D. if the URL includes @ symbol then the URL points to a phishing website. The authors conducted an empirical study to test the proposed program. The experiment was performed on 100 different URLs; here, the results classified 68% of them as legitimate while the others as malicious with an accuracy of 96% of phishing webpage detection. Authors in [6] propose an auto-updated whitelist to detect the phishing websites at the client side where it reports on the IP address and domain name. The normal behavior when applying the whitelists is to allow the entries involved and notify other non-included webpages whether it is deemed as legitimate or, otherwise, a phishing attack. Therefore, the idea for an auto-update is to test the webpages first by extracting the hyperlinks in the webpages and check them out, and then to decide whether to visit the website in the event that it is legitimate. This is in addition to adding it to the whitelist or, else, invoke a block action. This technique also allows for the detection of a DNS phishing attack and the Zero-hour phishing attack. In this, the system was designed using JAVA programming language, where the hyperlinks are extracted using Jsoup, where a check for the DNS domains the Google public DNS system is leveraged. The evaluation was conducted by applying an empirical study on the system by testing 1525 different webpages; the system offered an overall accuracy of 86.02%, implying a high degree of effectiveness. Authors in [7] propose a method known as Phishdentity which relies on favicon to determine the website’s identity with the help of Google image search engine which helps to use the images as a search query; here, there would be no need to using a database in addition to including the highest number of legitimate websites. Authors depend on certain features to compare the available websites with the authentic ones. The Second-Level Domain, which refers to the website name, Path in URL which refers to the location of the index.php file, the title and Snippet which refers to the text above the URL and the description under it, and the highlighted bold text (representing the text in the search results in addition to some others features) is used. These features have different weights based on their relative importance and effectiveness on the detection results then they are combined into some common frequency referred to as the final score. When this score is greater than a threshold then it is considered a phishing website. The empirical study
R. Al-Share et al.
was performed on 10000 different websites where the results show some promising indications. Authors in [8] propose a machine learning mechanism to detect malicious phishing emails; certain features were extracted from emails for detection purposes using Naive Bayes algorithm, followed by a set of ML algorithms as classifiers, where a final decision on the email class is examined using the decision tree algorithm. The system starts with an ML modeling which divides the input data to training and validation sets. This is used to develop the system model leveraging a Jupyter Framework Architecture. The authors built the model architecture and tested under a controlled environment. It was found that when evaluating the system model using a Random Forest algorithm a high degree of accuracy of 93. 54% was reached, and when two algorithms were combined the level of accuracy attained became 96.77%. Authors in [9] propose a deep learning framework called Conventional Neural Network Phishing Detection (CNNPD) to detect Email phishing attacks and classify them as legitimate sources or simply as phishing. In this work, this system was applied at the server side. To study and test the proposed model, the authors used different ham and phishing collected emails form PhishingCorpus and SpamAssasin public datasets. The header and body of the emails were extracted and the words were separated by spaces. Following that, deep learning is applied to the data. The deep learning first train and evaluate the input labeled dataset to build the model and get a better accuracy; then this model will be used to test the unclassified data set. CNN architecture is composed of four consecutive layers: Embedding layer, convolutional layer, pooling layer, and a fully connected layer. The email words were processed at each layer where the output of each layer is the input to the next. The output of the final layer is a classification of the email as phishing or benign. The proposed model was built in Python and the experimental test was conducted using a Google Collaboratory environment. The results yielded a 99.42% accuracy. Authors in [10] propose a hybrid methodology to detect occurrences of phishing emails and classify them using Support Vector Machine (SVM) Probabilistic Neural Network (PNN). The proposed model architecture starts with feature extraction, feature selection to determine which features to work on. The selected features undergo a Hybrid approach which consists of both a SVM classifier and a PNN classifier to determine whether the input email is phishing or legitimate. A dataset for phishing and legitimate emails was used. The system was tested and compared with other types of classes and with static and dynamic classifier methods and the results show better performance for the proposed hybrid model with an accuracy level of 98%.
20 Performance Evaluation of Online Website Safeguarding Tools Against Phishing Attacks; a Comparative Assessment
20.4
Phishing Detection Tools
Thousands of fake websites are found on the internet continuously and used to apply malicious scamming activities and steal sensitive information. The users are unaware of the real dangers that these websites can pose upon their personal lives. Many researchers have proposed several detection methods using machine learning and data mining algorithms to distinguish amongst phishing activities. The online website checkers are freely available on the internet to check for the validity of specific URLs, domains, and file contents. These usually use multiple security features that are tested against the websites to identify potentially malicious contents or behaviors. Website checkers could be used explicitly as an online service, like the Virus Total website and Norton Safe Web, or implicitly with the web browser, itself, which includes AVAST and AVG secure browsing. In this research, we have used Virus Total (VT) to test for the effectiveness of six scanning engines on a list of legitimate and malicious URLs collected from well-known anti-phishing frameworks, including PhishTank. VT is one of the well-known websites used to check files for viruses and scan for malicious URLs using 40+ antiviruses and online URL scanners [11]. We have used the best six scanning engines from VT in this study, including Sophos, Fortinet, Avira, CLEAN MX, Kaspersky, and BitDefender. 1. Sophos: Sophos works as a real-time antivirus and web filtering engine to defend against different security threats such as Malwares, ransomwares, AI threats, and web threats [12]. Sophos uses multiple features to detect imminent threats, including Live URL filtering, documents scanning, and reputation checking of the downloaded files. For URL filtering, as such, Sophos depends on a pre-built online database that includes a set of infected websites. It proactively detects maliciously downloaded contents and reviews the reputation of downloaded files by doing deep analysis on the file’s source and age. Sophos Endpoint Protection is a signature-based protection system presented by Sophos, and uses real-time threat intelligence to prevent unwanted malicious URLs, detects sudden changes, and controls flow of incoming traffic. Sophos works with different web browsers such as Google Chrome, Firefox, Edge, Internet Explorer and others. 2. Fortinet: A well-known antivirus system that protects against the newest viruses, spywares, and other malware types [13]. Its web filtering tools, including FortiGate, FortiClient, FortiCach, and FortiSandbox, all block phishing, hacked, or unwanted websites. Fortinet depends on automatic intelligent tools, periodic analysis, and continuous updates for web filtering. It uses a widely sorted and
3.
4. 5.
6.
165
categorized list to classify websites as accepted or blocked based on threat analysis. Fortinet has an intelligent defense system, which prevents malware downloads, controls access through a policy-based control system, and blocks access to malicious websites. FortiGuard blocks over 160,000 malicious websites every minute of the day. Its deep knowledge of the threat’s landscape and quick response enhances its protection efficiency [3, 4]. Avira: A security software that protects against different threats and cyberattacks such as web threats, identity theft, and other scamming activities [14]. Avira’s threat intelligence presents file reputation, file intelligence, web reputation, and domain categorization feeds. For the file reputation feed, specific information about the hashes, sizes, and formats of malicious files is delivered every 60 seconds. More detailed information regarding related certificates, associated exploits, attack vectors, malicious procedures, malware classes, and injection impacts on the systems are fed periodically by Avira’s File Intelligence Feed. For web reputation feeds, the URLs, timestamps, hashes, categories of malicious URLs are all fed and monitored. Avira blocks the known spamming, phishing, or malicious websites based on the information stored in its database. The URLs are classified into safe, malware, spam, PUA, phishing, or other types based on a realtime and site-specific query approach by using the web reputation API [5]. CLEAN MX: Is one of the sources that contain a large dataset of malicious verified URLs and malware sites [15]. Kaspersky: A cybersecurity company that uses deep threat intelligence and machine learning to present security solutions to businesses and users [16]. It protects small, medium, and large size enterprises from different cyber threats. Kaspersky Internet Security (KIS) is a product that offers protection against malwares, spams, and phishing activities. In the initial versions of this product, malware and spyware detection was done through Firewalls and Anti-Spy modules. Later, it added the sandboxing feature to its environment to run the applications on virtual machines. The URL adviser is a feature by KIS that enables the users to check websites and classify them into malicious and legitimate websites based on its database system. It color-marks these websites based on their severity level and degree of danger. BitDefender: A security software that uses advanced Artificial Intelligence (AI) to detect and block the newest types of threats [17]. An Anti-Phishing protection system is used to detect online scams associated with stealing of financial credentials like credit card information and passwords. Safe online banking enables users to make electronic payments safely by securing their transactions through a dedicated browser. This solution could prevent financial frauds while doing electronic shopping. It also
166
R. Al-Share et al.
uses the VPN to encrypt the internet traffic and protect banking information from sniffers. In Bitdefender, the URL Status Service identifies malicious, phishing, and fraudulent URLs and IP addresses in real-time. Bitdefender Cloud is used to scan websites without installing a scanning agent, which saves on memory and the associated processing of client’s devices.
20.5
the total number of legitimate websites, and F N denote the total number of legitimate websites correctly detected by the scanner as legitimate to the total number of legitimate websites.
Evaluation Procedure
In order to evaluate the performance of online scanners, a list of legitimate and malicious URLs was taken from multiple data sources, including PhishTank and SophosLabs. The chosen scanning engines were tested on the collected URL datasets, and a performance evaluation was done using the confusion matrix. A detailed description of the selected datasets and evaluation metrics are listed in the following subsections
20.5.1 Dataset Description We have applied the empirical study of URL scanning on two existing datasets collected from different sources. For malicious URLs, a dataset of 3500 websites was chosen from the sources [18, 19]. From source [18], datasets of malicious websites collected from the PhishTank data source, an open-source repository, where users submit suspicious websites continuously to the system since 2006. Additionally, we used a recent dataset that contains a list of phishing URLs collected during the period of COVID-19 pandemic by SophosLabs [19]. It has been noticed recently that the words “coronavirus” and “COVID19” were used frequently in various malicious domain names to launch phishing, malware, and spamming activities [20]. For legitimate URLs, a dataset of 3500 legitimate websites was chosen from the top 100K legitimate URLs collected in [18].
NP → P 100% NP
(20.1)
FP =
NL → P 100% NL
(20.2)
TN =
NL → L 100% NL
(20.3)
FN =
NP → L 100% NP
(20.4)
From a security perspective, having large false negatives is more dangerous than having false positives, which severely degrades the defense lines for businesses. In contrast, having large false positives might not be that dangerous, but it could become pretty dangerous when users ignore phishing alerts and consider them as normal events. However, in order to keep organizations more immune to any phishing activities, the detection tools should have low false positive and negative rates. Additionally, we have calculated the Precision, Recall, F1score, and Accuracy (ACC) for all scanning tools represented in Eqs. 20.5, 20.6, 20.7, and 20.8. Where P denote the ratio of correctly detected phishing websites to all instances detected as phishing, R denote the ratio of correctly detected phishing websites to all instances detected as phishing or legitimate from all malicious websites, F 1 Score denote the harmonic mean between Precision and Recall, and ACC denote the ratio of correctly detected phishing from all inputs.
20.5.2 Evaluation Metrics In order to evaluate the performance of online checkers, we first calculated the True positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN) values, which are represented in Eqs. 20.1, 20.2, 20.3, and 20.4. Where NP denote the total number of phishing pages, NL denote the total number of legitimate pages, T P : denote the number of malicious websites correctly detected by the scanner as malicious to the total number of malicious websites, F P denote the total number of legitimate websites wrongly detected by the scanner as malicious to the total number of legitimate websites, T N denote the total number of legitimate websites correctly detected by the scanner as legitimate to
TP =
P=
TP 100% TP + FP
(20.5)
R=
TP 100% TP + FN
(20.6)
2PR 100% P+R
(20.7)
F1 =
ACC =
TP + TN 100% TP + TN = FP = FN
(20.8)
20 Performance Evaluation of Online Website Safeguarding Tools Against Phishing Attacks; a Comparative Assessment
167
Table 20.1 Performance evaluation results for website checking tools Scanning engine Sophos Fortinet Avira ClLEAN MX Kaspersky BitDefender
20.6
Performance metric TP FP 98.46% 0% 59.74% 0.14% 77.94% 0% 61.29% 0.17% 95.63% 0% 82.60% 0%
TN 100% 99.86% 100% 99.83% 100% 100%
Testing Results and Discussion
Here, we present the results and outcomes of testing various URL scanning engines that we described earlier against the selected datasets. Table 20.1 presents the evaluation results of the six best URL scanners observed from VT website in terms of the evaluation metrics described previously. We can see from the results that Sophos outperformed other scanning engines with 99.23% accuracy and 100% precision values, which means that most of the websites were detected correctly by this scanning engine with no false positive values. Also, it achieved the highest recall value of 98.46, which indicates that the percentage of false negatives is very low compared with other scanning engines. Kaspersky comes in second place with an accuracy of 97.81%, followed by BitDefender (91.30%), Avira (88.97%), CLEAN MX (80.56%), and Fortinet (79.8%). All of the scanning engines achieved high precision values with more than 99% rate; Hence, almost all of them demonstrated very low false positive values. However, recall values fell in the range between 59% and 98%, depending on the false negative rates achieved by each engine. Compared with the empirical study made in [21] for detecting phishing websites, the scanning engines that we selected in our study achieved better detection performance in terms of detection accuracy. In [21], Anti-Phishing was the best detection tool for detecting malicious and legitimate websites with 94.32% accuracy, while Sophos achieved the best accuracy value with 99.23% in our study.
20.7
Conclusion and Future Work
Phishing attacks are among the most dangerous attacks that deceive users or employees in order to steal their personal information and impersonate their identities. Not only individuals are affected by these attacks, but also the organizations involved. Any data breach in their systems can affect their reputation and productivity in the market. Many organizations tend to use anti-phishing tools that identify malicious websites and spam emails to defend against phishing. In this empirical study, we evaluated the effectiveness of online
FN 1.54% 40.26% 22.06% 38.71% 4.37% 17.40%
Precision 100% 99.76% 100% 99.72% 100% 100%
Recall 98.46% 59.74% 77.94% 61.29% 95.63% 82.6%
F1-score 99.22% 74.73% 87.6% 75.92% 97.77% 90.47%
Accuracy 99.23% 79.80% 88.97% 80.56% 97.81% 91.30%
website checkers for detecting malicious URLs. Six scanning engines from the well-known security software packages were tested on 7000 malicious and legitimate URLs. The reputations of IP addresses, DNS servers, and website contents were used as parameters for detecting the malicious URLs. The results showed that Sophos achieved the best detection accuracy of 99.23% and 100% precision value. Also, all the tested engines achieved an accuracy above 50%.
References 1. K.L. Chiew, K.S.C. Yong, C.L. Tan, A survey of phishing attacks: Their types, vectors and technical approaches. Expert Syst. Appl. 106, 1–20 (2018) 2. S. Gupta, A. Singhal, A. Kapoor. A literature survey on social engineering attacks: Phishing attack, in 2016 International Conference on Computing, Communication and Automation (ICCCA). IEEE, 2016 3. B.B. Gupta, N.A.G. Arachchilage, K.E. Psannis, Defending against phishing attacks: Taxonomy of methods, current issues and future directions. Telecommun. Syst. 67(2), 247–267 (2018) 4. A.F. Alwaghid, N.I. Sarkar, Exploring malware behavior of webpages using machine learning technique: An empirical study. Electronics 9(6), 1033 (2020) 5. A.A. Ahmed, N.A. Abdullah, Real time detection of phishing websites, in 2016 IEEE 7th Annual Information Technology, Electronics and Mobile Communication Conference (IEM-CON). IEEE, 2016 6. A.K. Jain, B.B. Gupta, A novel approach to protect against phishing attacks at client side using auto-updated white-list. EURASIP J. Inf. Secur. 2016(1), 9 (2016) 7. K.L. Chiew et al., Leverage website favicon to detect phishing websites. Secur. Commun. Netw. 2018, 1 (2018) 8. B. Espinoza, et al., Phishing attack detection: A solution based on the typical machine learning modeling cycle, in 2019 International Conference on Computational Science and Computational Intelligence (CSCI). IEEE, 2019 9. R. Alotaibi, I. Al-Turaiki, F. Alakeel, Mitigating email phishing attacks using convolutional neural networks, in 2020 3rd International Conference on Computer Applications & Information Security (ICCAIS). IEEE, 2020 10. A. Kumar, J.M. Chatterjee, V.G. Díaz, A novel hybrid approach of SVM combined with NLP and probabilistic neural network for email phishing. Int. J. Electr. Comput. Eng. 10(1), 486 (2020) 11. VirusTotal (2012) [Online]. Available at https:// www.virustotal.com/gui/. Last accessed 2021 12. Sophos (1985) [Online]. Available at https://www.sophos.com/enus.aspx. Last accessed 2021 13. Fortinet (2009) [Online]. Available at https://www.fortinet.com/. Last accessed 2021
168 14. Avira (1986) [Online]. Available at https://www.avira.com/. Last accessed 2021 15. Clean-mx [Online]. Available at http://support.clean-mx.de/ cleanmx/viruses 16. Kaspersky (2005) [Online]. Available at https://me.kaspersky.com/ . Last accessed 2021 17. Bitdefender (2001) [Online]. Available at https:// www.bitdefender.com/. Last accessed 2021 18. ebubekirbbr, 2019, phishing url detection, GitHub. https:// github.com/ebubekirbbr/phishing. URL detection
R. Al-Share et al. 19. SophosLabs, 2020, covid-iocs/malicious URL, GitHub. https:// github.com/sophoslabs/covid-iocs/blob/master/malicious URL 20. S. Gallagher, A. Brandt, Facing Down the Myriad Threats Tied to COVID-19 [Online] (2020). Available at https://news.sophos.com/ en-us/2020/04/14/covidmalware/. Last accessed 2021 21. H. Sharma, E. Meenakshi, S.K. Bhatia, A comparative analysis and awareness survey of phishing detection tools, in 2017 2nd IEEE International Conference on Recent Trends in Electronics, Information & Communication Technology (RTEICT). IEEE, 2017
Part IV Blockchain Technology
Blockchain Based Trust for the Internet of Things: A Review
21
Dina Shehada, Maryam Amour, Suadad Muammar, and Amjad Gawanmeh
Abstract
Ensuring trust between Internet of Things (IoT) devices is crucial to ensure the quality and the functionality of the system. However, with the dynamism and distributed nature of IoT systems, finding a solution that not only provides trust among IoT systems but is also suitable to their nature of operation is considered a challenge. In recent years, Blockchain technology has attracted significant scientific interest in research areas such as IoT. A Blockchain is a distributed ledger capable of maintaining an immutable log of transactions happening in a network. Blockchain is seen as the missing link towards building a truly decentralized and secure environment for the IoT. This paper gives a taxonomy and a side by side comparison of the state of the art methods securing IoT systems with Blockchain technology. The taxonomy aims to evaluate the methods with respect to security functions, suitability to IoT, viability, main features, and limitations. Keywords
Evaluation · Social networks · Security attacks · QoS · Overhead · Privacy · Scalability · Taxonomy · Access control · Structure
devices to provide smart services to the users. It is noticeable that IoT is used by different industries to support a wide diversity of applications such as manufacturing, logistics, food industry, healthcare, utilities, smart homes, smart cities, smart grids, smart retail, etc.[1,2]. For example, in healthcare systems, IoT is used to assess and recognize patients afflicted with chronic illnesses [3]. IoT devices monitor and collect the physiological statuses of patients through sensors then send them to be analyzed to come up with the best diagnoses and inform physicians. Another example where IoT is highly beneficial is in supply chain management, where a product often consists of multiple parts provided by different manufacturers across countries. IoT is made up of interconnected “smart things” that are embedded with sensors, software, and other technologies to connect and exchange data with other devices and systems over the Internet. IoT enabling technologies are grouped into three categories; the technologies that enable “things” to acquire contextual information, process contextual information, and improve security and privacy. While the first two categories can be considered as functional building blocks required for building intelligence into things, the third category is a de facto requirement to severely eliminate penetration of IoT systems [2].
21.1.1 Problem Description and Motivation
21.1
Introduction
The Internet of Things (IoT) has become one of the most remarkable implemented technologies of the twenty-first century. The aim of IoT is to inter-connect all smart physical D. Shehada () · M. Amour · S. Muammar · A. Gawanmeh College of Engineering and IT, University of Dubai, Dubai, UAE e-mail: [email protected]; [email protected]; [email protected]; [email protected]
The intrinsic features of IoT cause several challenges mainly; decentralization, poor interoperability, privacy, and security vulnerabilities [1]. IoT systems are built with heterogeneous hardware and networking technologies, as a result, connecting them to the applications to extract and analyze large amounts of data is a complex task [4]. Moreover, trust management plays an important role in IoT to achieve trustworthy data collection, exchange, storage, and sharing. Therefore, there is a need to ensure trust among IoT devices. Many
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_21
171
172
D. Shehada et al.
Fig. 21.1 The structure of a Blockchain Previous Block
Block n Header
Block n+1 Header
Block n+2 Header
Hash of Block n-1 Header
Hash of Block n Header
Hash of Block n+1 Header
Block n Body
Block n+1 Body
Block n+2 Body
Data
Data
Data
researchers proposed mechanisms to ensure trust and security within different networks and applications. Some of these mechanisms include providing authentication, authorization, and trust evaluation between IoT devices. However, these trust mechanisms do not always fit the scale and diversity of the IoT, where there is no common root of trust and information extraction and processing is complex. A decentralized architecture is crucial to enable autonomy and flexibility in the system. Therefore, to provide the best possible services in any IoT application, there is a great demand to find an efficient solution that ensures trust and privacy, suitable to the nature of IoT environments and of course feasible. Many researchers are currently interested in Blockchain technology as it provides optimization elements to overcome trust and integrity issues. Due to its promising features, the Blockchain technology attracted many researchers to integrate it into many applications including IoT applications as it is considered a promising solution that can address the challenges of IoT [1]. In this paper, we surveyed the integration of Blockchain technology within IoT and proposed a taxonomy and provided a side by side comparison of the state of the art methods to assess Blockchain based techniques in IoT from trust perspective. The comparison assesses the techniques with respect to the security functions, suitability to IoT, viability, main features, and limitations. The rest of the paper is structured as follows. Section 21.2 discusses the background and related works in the literature. Section 21.3 explains the proposed taxonomy in detail. Section 21.4 explains how the proposed taxonomy is used to assess, analyze and compare the related works. Finally, conclusions and future work remarks are presented in Section 21.5.
21.2
Background and Related Works
21.2.1 Background Blockchain has gained more interest in IoT applications since it brings several opportunities and presents several solutions to address existing challenges in IoT [5]. Blockchain is
Next Block
a decentralized system that maintains data integrity using several consensus mechanisms [6]. It consists of several consecutively connected blocks where each block, except the first, points to its immediately previous block via an inverse reference, called the hash value of the parent block [1]. Blocks in a Blockchain containing details of transactions that have occurred within the network. The transaction information can be regarding token transfers occurring in a network, or any manner of data exchange. Each block is logically divided into two parts, namely, the header and the body. Transactions and important data are stored within the body of the block, while the header contains, among other fields, the identifier of the previous block [7]. The blocks are connected in a chain, similar to a linked list, as shown in Fig. 21.1. Transactions are verified and validated by entities in the network, then packaged into blocks and chained within the Blockchain network according to an established consensus mechanism. The fact that Blockchains use hash functions to chain blocks to ensure immutability, and encryption and digital signatures to secure information, makes Blockchain secure and tamperproof [6].
21.2.2 Related Works In this section, we review some of the most recent related works that have addressed security issues of IoT systems by incorporating the Blockchain technology. The paper in [8] proposed the use of Blockchain to secure and ensure integrity of datasets sharing with two conceptual Blockchain approaches. A Reference integrity Metric (RIM) is used to make sure that downloaded sets are trusted according to the metrics and a central hub is responsible for dataset distribution in addition to keeping track of system members. Both the system membership information i.e. ID, address, policy, etc. and the RIM are maintained using Blockchain. The membership information is publicly available to all members making privacy a major concern. Authors in [9] proposed a Blockchain Based Trust Management (BBTS) to maintain trust among network entities in smart cities applications such as the Internet of Vehicles
21
Blockchain Based Trust for the Internet of Things: A Review
(IoV). The proposed multi-tier architecture authenticates devices and manages trust calculation and distribution using a decentralized structure that is expected to minimize delays. Road Side Units (RSUs) are devices distributed in the smart city that also have great power and fast network connections. Platoons are groups of mobile nodes with limited power and network connections. Nodes within a platoon exchange messages about their experience and trust value of a specific node. After the messages are validated by them, the updated trust value is added to the platoon Blockchain. As nodes leave one platoon into another they share the previous platoon Blockchain with the new platoon’s RSU. Then, the RSU shares the new blocks with other RSUs to make sure that they all have a synchronized copy of the global Blockchain. Having nodes calculate the trust value provide a decentralized structure that ensures dynamism and flexibility of the system. However, this can cause overhead and a high computational cost on the system. Moreover, ensuring the consistency of the RSUs Blockchain is difficult due to the fact that the system is very dynamic. This also makes approximating the time needed for sharing and validating information difficult. In [10] the proposed system is made up of different manufacturing zones. Each zone contains devices and resources such as sensors, IoT devices, in addition to authentication and trust managers (servers). Authentication managers, authenticate and authorize devices, while trust managers are responsible for calculating the final trust values of devices. Moreover, a set of miners also exist to verify trust data, create the new block and broadcast it to be added to the Blockchain network. The system has a partially decentralized structure as trust is calculated by trust managers [11, 12], moreover, the Blockchain is only shared with verified devices that were granted permission, so a private Blockchain is used. Authors in [13] proposed a Blockchain based trust system for IoT to enhance the communication and dynamism in IoT systems. The system classifies IoT devices into three labels, lightweight, full, and coordination nodes depending on their computing capacity, storage, and connection time. Blockchain is used to manage all the unique IDs assigned to devices. Full nodes are responsible for assigning valid IDs to the devices while coordination nodes assign full nodes to lightweight nodes that newly joined the system and need to be verified and assigned a valid ID. According to the authors, it is difficult for attackers to fake IDs as the system maintains an ID management system through Blockchain. However, the authors did not explain how the global public Blockchain is maintained and synchronized to remain consistent. Authors in [14] proposed a Blockchain based system to create end to end trust between IoT devices without relying on the root of trust. The Blockchain uses cryptographic primitives to create a credit based Blockchain with reputation. IoT devices can perform transactions using that credit, and their ability to pay back their credit adds to their reputation. Trans-
173
actions are validated by consumer end devices by comparing them to some set terms of use (TERMs), then are publicly recorded in the Blockchain by a powerful back end server making the system have low latency. The proposal is expected to overcome the delays that usual Blockchains have [15, 16]. Authors in [6] proposed a partially decentralized authentication and access control system for IoT, based on multi agents systems (MASs) and Blockchain. The system is supposed to provide access control and secure communication between IoT devices, fogs and cloud. The system is made up of a hierarchical Blockchain architecture made up of four layers of Blockchain managers which are local, fog, core fog, and the cloud. Managers are equipped with high resources and responsible for controlling all communications and managing the Blockchain of their layer. The proposed framework in the latter study relies on multiple private Blockchain which improves privacy but affects scalability. Also, the proposal has not been implemented in a real-world environment [17, 18]. Authors in [19] proposed a decentralized trust management system for IoV using Blockchain. The Blockchain is used to ensure the security and integrity of trust values of the different IoT devices. Blockchain blocks are managed and saved at the fogs of the system where they are the only entities that can add new blocks to the chain. The system uses a public Blockchain that is distributed to all IoT devices and is updated only upon request. Distributed data are equipped with a timestamp and are updated after a proper timeout. Though this approach may reduce the overhead in the system, it also raises a concern about setting the proper value for timeout as it can affect the validity and trust of the whole system [20]. A decentralized interoperable trust framework for healthcare IoT (IoHT) is proposed in [3] to guarantee authentication and improve the trustworthiness between nodes and edges in the IoT system. Trusted zones are created after verifying the identity of the devices through cryptography and a public Blockchain. Once trusted zones are created, transactions requesting patient information are validated through a private ripple Blockchain hosted on the Health edge layer. Unlike the public Blockchains, the ripple Blockchain requires that the primary node among the group to grant access and permission to other nodes in order for them to be able to access information and write to the Blockchain [21]. An architecture for IoT combining fog computing and Blockchain based social networks is proposed in [22] to ensure the security of IoT environments. The proposal uses public Blockchain to make it easier for fogs and devices to record identities of devices and manage the relationship and trust between them. Whenever any device joins the system, it sends its information to the fogs where they are verified and recorded as a transaction. After that, the request is broadcast to miners (fogs) in the entire peer to peer network, once the
174
D. Shehada et al.
transaction is verified it is added to the public Blockchain. The system helps to establish trust in environments, by incorporating authentication and authorization mechanisms that improve the performance and security of the traditional IoT system [23, 24]. Another Blockchain based trust framework for IoT called IoT passport is proposed in [25]. The Blockchain based framework provides a trust management system among IoT devices. In addition to access control and permissions, all interactions among devices are signed by participants and recorded in the local Blockchain of the different trust domains. Later, a Blockchain node from each trust domain is chosen to synchronize and form the global public Blockchain. However, only a conceptual model is presented without any concrete implementation. Moreover, some issues such as how the global trust is formed or how Blockchain nodes are chosen are not discussed [24, 26]. The authors in [27] presented another Blockchain based trust evaluation for IoT devices, where reported histories stored in a Blockchain are used to compute the trust scores for devices. The trust score is evaluated using the proposed trust assessment where properties and capabilities set by the device manufacturer are used to find deviations and reports. Deviations define the behaviors that do not follow the capabilities and reports are the behavior feedback monitoring done by network controllers. Intelligent software
network based controllers store this information in a global Blockchain and then analyzers use it to compute the trust scores of devices and update the controllers to ensure that devices participating meet the trust level expectation. However, the proposal lacks concrete implementation and does not explain how the controllers and analyzers work together and use the information to evaluate the IoT devices [12]. In the next section, a taxonomy scheme is proposed to investigate the suitability of all the reviewed Blockchain solutions proposed in the literature.
21.3
Proposed Taxonomy Scheme
In this section, we propose a taxonomy scheme to assess the different reviewed papers and evaluate how suitable they are to IoT systems. The reviewed literature is analyzed thoroughly and classified based on the proposed taxonomy scheme. Figure 21.2, shows the proposed taxonomy and its classifications. The taxonomy evaluates the works in the literature according to five main categories; security functions, suitability to IoT environments, the proposed solution feasibility, its main features, and limitations. Next, we explain in detail each category and how it is used to assess the literature:
Taxonomy
Security
Suitability
Main Features
Feasibility
Limitations
Authentication
Structure
Implementation
Applications
Implementation
Access Control
Overhead
Testing & Validation
Integration
Requirements
Trust
Scalability
Privacy
Blockchain Management
Fig. 21.2 Proposed Taxonomy for classifications and analysis of existing solutions
21
Blockchain Based Trust for the Internet of Things: A Review
• The Security function: this item defines the different security requirements it provides. The main security functions provided by the works in the literature are authentication, access control, trust evaluation, and privacy. Privacy depends on the type of the Blockchain used whether it is public or private. Unlike public Blockchains, private Blockchains ensure the privacy of the end users by controlling which device can access the Blockchain. Moreover, although a public Blockchain brings numerous benefits, it can cause delays and overheads when utilizing a large amount of computational resources. • Suitability: this item evaluates how suitable the proposed solution is to IoT system, with respect to: – Structure: describes whether the Blockchain is centralized, decentralized, or partially decentralized. A centralized structure affects the availability of IoT systems and suffers from the problem of a single point failure therefore not recommended for IoT, while a decentralized structure is suitable to the nature and functionality of IoT. – Overhead: Identifies if the solution has a high overhead that can overwhelm the system and cause delays. This is why it is important to consider and evaluate the amount of computations and communication overhead caused by the different proposed solutions. – Scalability: evaluating how easy it is for the system to expand. The structure of the proposed solution has a great impact on the scalability. For example, relying on the cloud makes it difficult to expand the system, while having distributed fogs responsible for registering and monitoring devices makes it easier and faster. – Blockchain management: defines which entity/ entities have the responsibility for managing the Blockchain, its storage, and its distribution. This feature is also important as it has a direct impact on the system overhead and security and therefore on the suitability to IoT systems. • Feasibility: this item identifies the feasibility of implementing and verifying the proposed solution. – Implementation: shows the feasibility of the solution in terms of implementation and deployment through practical Blockchain. – Testing and validation: shows the validation of the solution throughout simulation or prototype. • Features: listing the main features that make the proposed solution unique. – Application: shows the type of applications proposed by the approach. – Integration: shows integration with other architectures such as fog and edge computing. • Limitations: Listing any limitations that govern the proposed solution.
175
– Implementation: shows limitations related to implementing the solution. – Requirements: shows requirements and assumptions under which the proposed solution works. The proposed taxonomy is used to assess [3, 6, 8–10, 13, 14, 19, 22, 25, 27] from the literature. Details of the assessment are explained in the next section.
21.4
Assessment of Blockchain Based Trust Approaches in IoT
Most of the proposed works use Blockchain to save information and policies to provide access control such as [8,13,14]. Some provide both authentication and access control such as [3, 6, 22, 25]. While other works also save the computed trust score for entities in the Blockchain to provide trust calculation and authentication [9, 10] or access control [19, 27]. All the proposed works have a public Blockchain scheme except [6, 10] that use a private permissioned Blockchain to ensure privacy. While [3] proposed a solution that uses both public Blockchain in addition to a private ripple Blockchain. The public Blockchain is used to ensure authentication to devices, while the ripple is used to enforce authorization and access control. The framework in [8], has a centralized structure as it relies on a single central hub to manage the Blockchain. Furthermore, the frameworks in [9, 13, 19, 22, 25, 27] have a fully decentralized structure as they use distributed fogs, nodes or devices to complete the Blockchain management and distribution. On the other hand, the rest of the reviewed frameworks [3, 6, 10, 14] are considered to have a partially decentralized structure due to their reliance on multiple but specific entities to manage the Blockchain such as agents, edge devices or servers. When it comes to the overhead, it is important to have a lightweight trust solution to suit the nature of the IoT systems and make their expansion easier. The frameworks proposed in [6,10,14,19,22] managed to provide that by giving the system access to powerful devices or servers that can handle any computations. On the other hand, [3, 25, 27] are considered to have a relatively higher overhead. In [3] the nature of the application requires a lot of communication between devices as any request or permission need to be managed by the primary nodes and the Health edge. In [25], heavy communications need to take place between the nodes to synchronize the Blockchain between the different trust domains. Additionally, [27], relies on controllers and analyzers to evaluate trust and generate reports. The rest are considered to have an average overhead as they either rely on broadcasting blocks to IoT devices for verification [8] or give them the responsibility for creating, sharing,
176
D. Shehada et al.
and managing the Blockchain [9] that can be overwhelming to such devices with limited power and storage. Also, in [13] the process of nodes classification, and assignment of coordination nodes to other nodes can be overwhelming on the system. As for scalability, [3, 9, 10, 13, 14, 19, 22, 25, 27] are considered scalable solutions that are easy to expand to add more IoT devices due to the easy deployment and management of the Blockchain structure. Banerjee et al. [8], Lahbib et al. [10] have medium scalability due to the centralized structure. The framework in [6], is not considered to be scalable, as its multi layer scheme, in addition to the private Blockchain used and the issue of the big header size in their proposed framework make the system expandability complicated. According to the proposed work in the literature the Blockchain management is done by different entities. For example, in [19, 22] the Blockchain management is conducted by the Fogs. Fogs add blocks to the Blockchain and are responsible for its distribution and storage. In [8] the Central Hub does most of the management while in [9], both nodes
and RSUs do. Full nodes manage and store the Blockchain in [13]. A powerful service provider back end server manages the Blockchain in [14]. In [6] each layer manages its Blockchain by a specific agent called Blockchain manager. In some cases, the Blockchain management is conducted by more than one entity as in the case of [3] where both the Health edge and the primary nodes carry out the task together. In [25, 27], Blockchain storage and management is done by Blockchain nodes and controllers, respectively. While in [10] miners and the trust server take care of that. The work in [19] was implemented through Hyperledger Fabric tool. Also [3, 10] conducted a simulation using NS3 simulation and C# & JSON - LD symantic tools. While the work in [22], was validated by building a prototype using SE Linux. Table 21.1 summarizes the assessment of the related works based on the proposed comparison scheme. Moreover, the main features, challenges, and limitations s are also highlighted in Table 21.2.
Table 21.1 Assessment of Blockchain based trust approaches in IoT Framework
Security Function
Banerjee et al. Access control [8] Kandah et al. [9] Authentication and trust score evaluation Lahbib et al. [10] Authentication and trust score evaluation Qiu et al. [13] Access control Di Pietro et al. Access control [14]
Algarni et al. [6]
Authentication and Access control Cinque et al. [19] Access control and trust score evaluation Abou-Nassar et Authentication al. [3] and Access control Zhu and Badr Authentication [22] and Access control Tang et al. [25] Authentication and access control Boussard et al. Access control [27] and trust score evaluation
Blockchain type
Suitability to IoT Structure Centralized
Overhead Scalability Blockchain management Average Medium Central Hub
Feasibility Implementation and testing No
Public Public
Decentralized
Average
Private
High
Nodes & RSU
No
Partially decentralized Low
Medium
Trust servers and miners
Simulation
Public Public
Decentralized Average Partially decentralized Low
High High
No No
Private
Partially decentralized Low
Low
Public
Decentralized
Low
High
Full nodes Service provider backend server Blockchain manager at each layer Fogs
Public and Private Partially decentralized High
High
Simulation
Public
Decentralized
Low
High
Health Edge and primary nodes Fogs
Public
Decentralized
High
High
Blockchain nodes
No
Public
Decentralized
High
High
Controllers
No
No
Implementation
Prototype
21
Blockchain Based Trust for the Internet of Things: A Review
177
Table 21.2 Summary of features, challenges, and limitations of proposed Blockchain based trust approaches in IoT Framework Banerjee et al. [8]
Features
Challenges and limitations
o Secure sharing of IoT datasets o Ensures dataset integrity
o Deciding time of data sharing o Cannot avoid privacy violations
Kandah et al. [9]
o Provides trust among IoT entities in smart cities o o Single bad mouther has a minimal effect o o Protection from colluding attacks if less than 33% of o the vehicles are malicious o
Lahbib et al. [10]
Assumes that miners can all be trusted o Blockchain based trust architecture o Collects trust evidences and gives trust scores for IoT devices o Resilient to Tamperproof, Bad Mouthing, On Off and Ballot Stuffing attacks
Qiu et al. [13]
o Dynamic and scalable architecture for IoT o How public Blockchain is formed by the different full o Classifies nodes based on their connectivity, computing nodes and storage o Trusting full nodes
Di Pietro et al. [14]
o Creates endtrust between devices without relying on o Does not specify how terms are published the root of trust o Privacy issues o Credit based Blockchain o Low latency
Algarni et al. [6]
Provides secure access control based on MAS and Blockchain
o Delays caused due to the big header size in the proposed Blockchain structure o Difficult to implement
Cinque et al. [19]
o Secures the trust management of IoV o Ensures security and integrity of trust value score
o Setting a suitable timeout time to update value o Does not provide protection against attacks
Privacy issue Trustworthiness of connected devices need to be evaluated Ensuring consistency among platoons by RSU Approximating time of message sharing
Abou-Nassar et al. [3] o Interoperable trust framework using ripple Blockchain o Relies on primary nodes to allow access to Blockchain o Provides privacy while not limiting access to a single information person o High overhead Zhu and Badr [22]
o o o o
Hybrid architecture Fog computing enabled Blockchain based social networks Distributed public ledgers for recording identities
Privacy issues
Tang et al. [25]
o o o o
Blockchain based trust frameworks Cross platform collaboration between IoT devices Interactions are recorded on the Blockchain Interactions are utilized in an incentive plan
o Performance of contemporary Blockchain infrastructure is limited o There is a need for other mechanisms to ensure synchronization of Blockchains
Boussard et al. [27]
o Blockchain based trust evaluation o Assumes a baseline of expected network behaviour o Reported history is analyzed using simple risk assess- o Can be used to monitor and report IoT devices ment o Automated (dis)connections can be perceived as unjuso Provides a trust score that is calculated using capabilitified and authoritative ties, deviations and reports
178
21.5
D. Shehada et al.
Conclusion
The issues related to the security and privacy of the IoT systems are immense and require careful consideration. In order to effectively counteract these security and trust issues, this research proposes a taxonomy derived from the synergy of Blockchain and trust in IoT. The taxonomy is used to assess the existing Blockchain based trust approaches in the literature. A comprehensive review of the approaches along with a detailed analysis about their security features, suitability to decentralized IoT communities, and feasibility is provided through a side by side comparison. We also highlight the main features and limitations to help the reader in effectively evaluating the different proposed approaches.
References 1. H.-N. Dai, Z. Zheng, Y. Zhang, Blockchain for internet of things: a survey. IEEE Int. Things J. 6(5), 8076–8094 (2019) 2. O. Vermesan, P. Friess, et al., Internet of Things-from Research and Innovation to Market Deployment, vol. 29 (River Publishers, Aalborg, 2014) 3. E.M. Abou-Nassar, A.M. Iliyasu, P.M. El-Kafrawy, O.-Y. Song, A.K. Bashir, A.A. Abd El-Latif, Ditrust chain: towards blockchainbased trust models for sustainable healthcare iot systems. IEEE Access 8,111223–111238 (2020) 4. C. Sobin, A survey on architecture, protocols and challenges in IoT. Wireless Personal Commun. 112(3), 1383–1429 (2020) 5. M.A. Uddin, A. Stranieri, I. Gondal, V. Balasubramanian, A survey on the adoption of blockchain in IoT: challenges and solutions. Blockchain: Res. Appl. 2, 100006 (2021) 6. S. Algarni, F. Eassa, K. Almarhabi, A. Almalaise, E. Albassam, K. Alsubhi, M. Yamin, Blockchain-based secured access control in an IoT system. Appl. Sci. 11(4), 1772 (2021) 7. M.S. Ali, M. Vecchio, M. Pincheira, K. Dolui, F. Antonelli, M.H. Rehmani, Applications of blockchains in the internet of things: a comprehensive survey. IEEE Commun. Surv. Tutorials 21(2), 1676–1717 (2018) 8. M. Banerjee, J. Lee, K.-K.R. Choo, A blockchain future for internet of things security: a position paper. Digital Commun. Netw. 4(3), 149–160 (2018). [Online]. Available: https://www. sciencedirect.com/science/article/pii/S2352864817302900 9. F. Kandah, B. Huber, A. Skjellum, A. Altarawneh, A blockchainbased trust management approach for connected autonomous vehicles in smart cities, in 2019 IEEE 9th Annual Computing and Communication Workshop and Conference (CCWC) (2019), pp. 0544–0549 10. A. Lahbib, K. Toumi, A. Laouiti, A. Laube, S. Martin, Blockchain based trust management mechanism for IoT, in 2019 IEEE Wireless Communications and Networking Conference (WCNC) (IEEE, Piscataway, 2019), pp. 1–8 11. A. Sharma, E.S. Pilli, A.P. Mazumdar, P. Gera, Towards trustworthy internet of things: a survey on trust management applications and schemes. Comput. Commun. 160, 475–493 (2020)
12. B. Shala, U. Trick, A. Lehmann, B. Ghita, S. Shiaeles, Blockchain and trust for secure, end-user-based and decentralized IoT service provision. IEEE Access 8, 119961–119979 (2020) 13. H. Qiu, M. Qiu, G. Memmi, Z. Ming, M. Liu, A dynamic scalable blockchain based communication architecture for IoT, in International Conference on Smart Blockchain (Springer, Berlin, 2018), pp. 159–166 14. R. Di Pietro, X. Salleras, M. Signorini, E. Waisbard, A blockchainbased trust system for the internet of things, in Proceedings of the 23nd ACM on Symposium on Access Control Models and Technologies (2018), pp. 77–83 15. M.S. Ali, M. Vecchio, M. Pincheira, K. Dolui, F. Antonelli, M.H. Rehmani, Applications of blockchains in the internet of things: a comprehensive survey. IEEE Commun. Surv. Tutorials 21(2), 1676–1717 (2018) 16. J. Huang, L. Kong, G. Chen, M.-Y. Wu, X. Liu, P. Zeng, Towards secure industrial IoT: blockchain system with credit-based consensus mechanism. IEEE Trans. Ind. Inf. 15(6), 3680–3689 (2019) 17. B. Alouffi, M. Hasnain, A. Alharbi, W. Alosaimi, H. Alyami, M. Ayaz, A systematic literature review on cloud computing security: threats and mitigation strategies. IEEE Access 9, 57792–57807 (2021) 18. S. Pal, A. Dorri, R. Jurdak, Blockchain for IoT access control: Recent trends and future research directions (2021). arXiv Preprint arXiv:2106.04808 19. M. Cinque, C. Esposito, S. Russo, O. Tamburis, Blockchainempowered decentralised trust management for the internet of vehicles security. Comput. Electri. Eng. 86, 106722 (2020) 20. C. Esposito, M. Ficco, B.B. Gupta, Blockchain-based authentication and authorization for smart city applications. Inf. Proc. Manag. 58(2), 102468 (2021) 21. M.A. Rahman, M.S. Hossain, M.S. Islam, N.A. Alrajeh, G. Muhammad, Secure and provenance enhanced internet of health things framework: a blockchain managed federated learning approach. IEEE Access 8, 205071–205087 (2020) 22. X. Zhu, Y. Badr, Fog computing security architecture for the internet of things using blockchain-based social networks, in 2018 IEEE International Conference on Internet of Things (iThings) and IEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart Data (SmartData) (IEEE, Piscataway, 2018), pp. 1361–1366 23. H. Wang, L. Wang, Z. Zhou, X. Tao, G. Pau, F. Arena, Blockchainbased resource allocation model in fog computing. Appl. Sci. 9(24), 5538 (2019) 24. P. Cui, U. Guin, A. Skjellum, D. Umphress, Blockchain in IoT: current trends, challenges, and future roadmap. J. Hardw. Syst. Security 3(4), 338–364 (2019) 25. B. Tang, H. Kang, J. Fan, Q. Li, R. Sandhu, Iot passport: A blockchain-based trust framework for collaborative internet-ofthings, in Proceedings of the 24th ACM Symposium on Access Control Models and Technologies (2019), pp. 83–92 26. C. Esposito, O. Tamburis, X. Su, C. Choi, Robust decentralised trust management for the internet of things by using game theory. Inf. Process. Manag. 57(6), 102308 (2020) 27. M. Boussard, S. Papillon, P. Peloso, M. Signorini, E. Waisbard, STewARD: SDN and blockchain-based trust evaluation for automated risk management on IoT devices, in IEEE INFOCOM 2019-IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS) (IEEE, Piscataway, 2019), pp. 841–846
The Use of Blockchain Technology in Electronic Health Record Management: An Analysis of State of the Art and Practice
22
Henrique Couto, André Araújo, Rendrikson Soares, and Gabriel Rodrigues
Abstract
Driven by the need to offer digital solutions to the population, the healthcare sector requires computational solutions with features of security, immutability and traceability for data transactions on the Electronic Health Record (EHR). An EHR is defined as a repository of healthcare information stored and transmitted in a secure and accessible way by authorized users. To address this important area of research, this work investigated state of the art practice and studies that addressed the development and validation of computational solutions with Blockchain technology applied to the following areas of an EHR lifecycle: (i) modeling and standardization (ii) data storage techniques, (iii) standards for data interoperability, and (iv) data retrieval and visualization solutions. Based on the results found, this study presents an analysis of the main advances and opportunities identified in the use of Blockchain technology in the development of healthcare applications. Keywords
Blockchain · Electronic health record · Health applications · Distributed databases · Review study · H. Couto () Computing Institute, Federal University of Alagoas, Maceió, Brazil e-mail: [email protected] A. Araújo Computing Institute, Federal University of Alagoas, Maceió, Brazil Advanced Studies in Data Science and Software Engineering, Federal University of Alagoas, Penedo, Brazil e-mail: [email protected] R. Soares · G. Rodrigues Advanced Studies in Data Science and Software Engineering, Federal University of Alagoas, Penedo, Brazil e-mail: [email protected]; [email protected]
Data modeling · Data storage techniques · Data interoperability · Data security · Smart contracts
22.1
Introduction
The health sector has undergone major transformations over the past decade due to the adoption of information and communication technologies (ICT) in clinical patient care, and to the popularization of the access to Electronic Health Record (EHR) information [1]. Commonly called eHealth, this digital transformation process aims to improve healthcare practice and reduce operational costs involved in patient care [2]. The insertion of digital business models in the healthcare sector is driven by the demands of society and the global economy for functional and safe computing solutions that positively impact the population’s quality of life. In the ecosystem in which health institutions are inserted, several software systems are used to process health related data. These software systems may include an EHR, telemedicine applications, examination and diagnostic platforms, as well as mobile applications that manage and monitor an individual’s clinical data. Due to the diversity of software applications present in healthcare, this sector requires solutions for the modeling and standardizing EHR data, data interoperability, and ensuring the integrity and authenticity of patient data shared in digital media [3, 4]. Advancements in computer technologies have made it easy to share data to quickly access information and provide convenience in carrying out healthcare services [5]. However, the volume of data processed by such software systems makes the authenticity of data transferred between healthcare organizations and end users an important and relevant research question within the scientific community [6]. Nowadays, accessing and sharing EHR data has become
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_22
179
180
a common practice for two reasons. First, the end user, also known as e-patient, needs to access their clinical data history and grant access rights to their data to third parties if needed. Second, for administrative reasons, healthcare institutions need to exchange data with other organizations and governmental regulatory agencies. Blockchain technology has been studied and applied in Academia and in the software industry as an alternative to guarantee the authenticity, traceability and security of online data transactions [7]. It consists of a decentralized and encrypted network that immutably certifies and stores all transactional information between the parties involved [8]. Blockchain has been used in the most diverse areas of knowledge [8], and in the healthcare sector, numerous studies have used its foundations and characteristics to provide solutions for the ecosystem of healthcare organizations [9– 11]. Although this topic has been debated and applied in this sector, some Blockchain-related research issues that lack advances and investigation are reported in the state of the art, such as the generation of Blockchain transactions from EHR data, and the interoperability of clinical data between healthcare organizations [12]. In order to analyze how Blockchain technology has been applied in the healthcare sector, this paper investigates how the characteristics of decentralized storage, which promote security, immutability and traceability, have been applied in the management of an EHR lifecycle. To achieve this goal, the following steps were carried out: (i) a search strategy was defined using Blockchain and EHR related terms, (ii) the search was carried out in the main repositories covering this field, (iii) the inclusion criterias were the development and validation of computational solutions with Blockchain technology applied to data modeling and standardization, storage techniques, standards for data interoperability, and data visualization solutions. Finally, (iv) the search covered studies published between 2016 and 2021. Based on the results found, this study presents an analysis of the main advances and opportunities identified in the use of Blockchain technology aimed at the development of healthcare applications. The remaining sections of this article are organized as follows. Section 22.2 describes the basic concepts needed to understand Blockchain technology. Section 22.3 presents the main studies identified in state of the art and practice, while Sect. 22.4 discusses the main advances found and provides an analysis of research opportunities, while final considerations and indications for future work are found in Sect. 22.5.
H. Couto et al.
22.2
Background
Blockchain technology is a decentralized database that differs from others by its data storage structure. In Blockchain, data is stored in sets called blocks which are persisted in transactions, where the network nodes are responsible for authorizing the entry of these blocks into the Blockchain [13]. Blocks are arranged in a linked list, where each block has a reference to the previous block and only one new block can be added at a time. The beginning of the chain of blocks is determined by the genesis block, which stores the information that must be present in other blocks [13]. For a transaction to be carried out, the nodes that make up the network must authorize the entry of the respective blocks. In this sense, there is a consensus process between the nodes so that new blocks are added to the network. Different types of consensus algorithms can be used, among which the Proof-of-Work (PoW), the Proof-of-Stake (PoS) and the Proof-of-Burn (PoB) stand out. In PoW, each node must solve a mathematical challenge in order to find a value for the hash code belonging to the block through the process of trial and error (or brute force) [14]. PoS is characterized by conducting a draw that determines which nodes will be responsible for authorizing a new block in the Blockchain network. In this case, users who wish to participate in the process need to deposit a certain amount of coins on the network as proof of their participation [15]. In PoB, nodes must burn some coins as proof of commitment to the network by obtaining the right to mine [16]. As the coin burning process represents virtual mining power, the more coins a user burns in favor of the system, the more mining power they have, and therefore, the better their chances of being chosen as the next validator on the block.
22.3
Blockchain-Based EHR Management Analysis
This section presents the study carried out on the use of Blockchain technology in an EHR lifecycle, and is organized as follows. In Sect. 22.3.1, the advances identified in the state of the art in the research area of this study are discussed, while in Sect. 22.3.2, the computational solutions created by the software industry and that are in use in the healthcare sector are presented.
22 The Use of Blockchain Technology in Electronic Health Record Management: An Analysis of State of the Art…
22.3.1 An Analysis of the State of the Art The development and maintenance process of a Health Information System (HIS) is characterized, among other things, by the need for a database project. A database design represents an important phase in the lifecycle of a software system, as it allows to identify and model EHR data requirements. In addition, it allows the design of an architecture with components that offer the best way to store and retrieve data. In this sense, the use of Blockchain technology in healthcare applications is justified by its ability to provide greater reliability in data storage. Through Blockchain, it is possible to verify the integrity and guarantee the immutability of all records stored on the network. In order to investigate how Blockchain technology is being applied in the healthcare sector, studies that address the development and validation of computational solutions applied to the following areas of an EHR were considered in this study: (i) data modeling and standardization, (ii) data storage techniques, (iii) standards for data interoperability, and (iv) data retrieval and visualization solutions. The criteria defined above were chosen because they deal directly with an EHR lifecycle, which starts with the modeling phase, and goes on with data visualization and software applications. Recent research involving Blockchain technology has focused on controlling access to data processed by healthcare organizations and integrating multiple Blockchain networks [17]. Gordon and Catalini [12] report that healthcare interoperability generally focuses on data exchange between commercial entities, however, solutions aimed at interoperability and data access for patients need to be developed. Pandey and Litoriya [18] propose a decentralized data storage software architecture using Blockchain technology and a computational service that allows healthcare organizations to access data without compromising the security of information transferred between stakeholders. Likewise, Zhang et al. [19] address the traffic of patient data between users and medical applications securely through Blockchain technology. In [6], a tool is presented that aims to develop an organizational governance process to achieve the semantic interoperability of clinical models in healthcare. In [20], a cloud computing architecture to connect hospitals and provide APIs for information consultation is discussed. In the same line of research, Huang et al. [21] present a healthcare system in which records are encrypted and stored on a Blockchain network so that only patients and their healthcare providers can view or update them. The system uses a permit management system through smart contracts where patients control which institutions can access the historical record of their clinical data. Using the same approach, a hybrid architecture using Blockchain and Edge nodes is developed in [22] to
181
facilitate EHR access control, and a software system that uses a Cross-Blockchain platform for data sharing and privacy preservation of patient information is discussed in [23]. The use of Blockchain technology to store EHR data in cloud services requires IT resources with great processing power and high bandwidth Internet. In order to minimize the dependence on large computational resources, the solution developed in [24] uses a fog architecture to process and store data from patients who need clinical care. Following this same line of research, Nagasubramanian et al. [25] developed a framework based on a cloud architecture to ensure the integrity of the EHR. In conjunction with Blockchain technology, the framework makes use of an encryption approach called Keyless Signature Infrastructure (KSI), which is based on the hash function and provides security for the system that uses it. According to the results presented, the response time of the proposed framework with Blockchain technology is almost 50% lower than conventional techniques and the storage cost is about 20% lower for the Blockchain system compared to existing techniques. In order to minimize the use of computational resources to process Blockchain transactions, Chen et al. [26] propose a searchable Blockchainbased encryption scheme for EHR. In this solution, only one index is stored in Blockchain technology, which allows data to be searched using logical expressions. In conjunction with Blockchain technology, healthcare standards such as OpenEHR and HL7 archetypes were used to create software applications aimed at managing EHR data. The openEHR standard addresses issues on how to standardize EHR data attributes, constraints and terminology [27], while the HL7 standard provides a set of specifications aimed at standardizing the exchange of information between software systems [28]. In this sense, the OpenEHR standard was used in [29] to create a distributed software architecture in which patients can record and consult their clinical care history from any device. The HL7 standard was used in [30] to create a Blockchain-based, globally integrated health record sharing architecture. The proposed architecture allows the patient to grant access to their clinical data and find which entities have accessed their EHR. In turn, Alamri and Margaria [31] propose a framework that makes use of HL7 to guarantee the compatibility of data exchange between different systems, Internet of Things (IoT) and EHRs, in a simple and understandable approach that saves time and costs. To make the application of the investigated works easier to understand, Fig. 22.1 classifies the area of an EHR lifecycle for which each research solution was developed. The analysis of the advances made and the research opportunities identified are discussed in Sect. 22.4.
182
H. Couto et al.
Research Roehrs et. al, 2021 [17]
EHR Lifecycle M
S
ᅛ
I ᅛ ᅛ
Gordon and Catalini (2018) [12] ᅛ
Pandey and Litoriya (2020) [18]
ᅛ
Zhang et al. (2018) [19]
ᅛ
Carmen et al. (2016) [6]
ᅛ ᅛ
Dubovitskaya et al. (2017) [20] Huang et al. (2021) [21]
ᅛ
ᅛ
ᅛ
ᅛ ᅛ
Islam et al. (2019) [24]
ᅛ
Nagasubramanian et al. (2020) [25]
ᅛ
Chen et al. (2019) [26]
ᅛ
Roehrs et al. (2019) [29]
ᅛ
ᅛ
Guo et al. (2019) [22] Wang and He (2021) [23]
V
ᅛ
ᅛ
Abdeen et al. (2019) [30] Alamri and Margaria (2021) [31]
ᅛ ᅛ ᅛ
M - Modeling || S - Storage || I - Interoperability || V - Visualization
Fig. 22.1 Categorization of investigated works
22.3.2 An Analysis of the State of the Practice Driven by the need to offer digital solutions to the population, the healthcare sector requires software systems with features of security, immutability and traceability for processed transactions. In this context, tools and software platforms using Blockchain technology that are directly related to EHR data management were investigated in the State of practice. BurstIQ [32] is an ecosystem that integrates Blockchain technology and machine learning to allow data from different sources to be stored in a single data repository. Here, users have full control over their clinical data and can share or use it in exchange for access to products and services of interest to them. HealthNet [33] is a health data sharing platform focused on tracking clinical data, where it is possible to trace everything from prescription drugs to insurance claims and coordination of care between providers for patients. MediLedger [34] is a permissioned Blockchain platform developed for the pharmaceutical industry based on open standards and specifications (e.g., HL7, OpenEHR). Its goal is to enable cost savings for healthcare organizations and provide interoperability and security features. Corel Health [35] is an iterative, interoperable, accessible, secure, and scalable health ecosystem that ensures data
security and accessibility through the use of smart contracts and Blockchain. Its goals include solving data warehousing issues, the incongruity of legacy databases, the challenges of analyzing unstructured data, high administrative costs, and lack of data security. In the same line of tools, Medicalchain [36] is a platform that uses Blockchain technology to store EHR data and allow users to share it with healthcare professionals and organizations of their choice. Furthermore, it is possible to create different health applications that use Medicalchain to perform the storage of users’ health data, creating an ecosystem where EHR data can be shared between different applications. Finally, there are also initiatives such as Nebula Genomics [37], a big data company for the sequencing and health of the human genome that uses Blockchain to store data. Its goal is to build a reliable genomic data marketplace for consumers, researchers and the medical community.
22.4
Discussion
Figure 22.2 categorizes the works identified in the state of the art, observing their application of Blockchain technology in the following areas of the EHR: (i) data modeling and standardization, (ii) data storage techniques; (iii) standards for data interoperability; and (iv) data retrieval and visualization solutions. As shown in Fig. 22.2, most of the works mentioned in this study use Blockchain technology to provide data interoperability features between healthcare organizations. The large number of studies in the area of interoperability is justified by the high traffic of data exchanged between stakeholders involved in the health ecosystem, that is, private and public organizations, government agencies, health regulatory agencies and the patients themselves. All information exchanged online requires computational resources with characteristics of access control, data integrity and immutability, resources that are provided by Blockchain technology. For data exchange to take place between healthcare industry stakeholders, storage solutions need to be developed to manage Blockchain transactions. As shown in Fig. 22.2, several studies have worked in this field of research and developed different solutions to store EHR data using Blockchain technology. However, some works only store an EHR index on the Blockchain network, while other works store all the data that characterize a patient’s EHR. Also in this field of research, it was observed that data storage with Blockchain technology commonly occurs in two ways. The first is to use platforms that are capable of creating decentralized software applications (DApps) that work on a peer-to-peer network. DApps are open source and operate autonomously and independently of central mediators [38]. Adaptations and changes to DApps only occur through
22 The Use of Blockchain Technology in Electronic Health Record Management: An Analysis of State of the Art…
183
Fig. 22.2 Analysis of the investigated works by area
consensus between the parties involved. A decentralized software application interacts with Blockchain technology through smart contracts, which are self-executable protocols coded directly in the Blockchain that establish a set of rules for the interaction to occur in an autonomous and transparent way [39]. The other way is to use a DBMS that offers Blockchain technology capabilities. In this case, the application’s business rules are implemented in the programming language chosen in the project, and the data persistence service is specified taking into account the data structures, software architecture and how transactions are implemented by the DBMS. Once the analysis of the solutions developed for the areas of modeling and data visualization was concluded, the following points were identified. As seen in Fig. 22.2, few studies addressed the development of software applications with data visualization capabilities stored in Blockchain technology through graphical user interfaces. This highlights the need to develop software systems that address data visualization aspects and the construction of graphical user interfaces taking into account usability concepts for different types of applications (e.g., web, mobile). Finally, few works were identified in the area of modeling and standardizing EHR data for creating Blockchain transactions. Of the four areas of EHR management analyzed in this research, the area of
modeling and standardizing EHR data is more likely to be seen as crucial by healthcare organizations, since it directly deals with the construction of Blockchain transactions from data processed by software systems in a health organization. The importance of this field of research is justified by the heterogeneity of data found in the healthcare sector, the complexity of managing data processed by various software systems, and the lack of uniformity in EHR data, caused by the absence of health standards (e.g. openEHR archetypes, ICD-10, Snomed) in software system development. The challenges reported above show that studies addressing the standardization of EHR data must be carried out. In particular, to support healthcare organizations in building Blockchainbased computational solutions to offer features of security and immutability for legacy EHR data. The tools and platforms analyzed in the state of practice represent an important contribution from the software industry to the healthcare ecosystem. Business models are being transformed through Blockchain technology, which has allowed security in the storage of EHR data, access granted by patients, and the traceability of data transactions. However, although the tools and platforms presented in this study represent an important advance, it is reported in the software industry that the development of computational environments for the creation of DApps and the insertion of
184
H. Couto et al.
Blockchain functionalities in DBMS require further research and development.
22.5
Final Considerations
This work presents a systematic review of research addressing the development and validation of computational solutions applied to the following areas of the Electronic Health Record (EHR): (i) data modeling and standardization, (ii) data storage techniques; (iii) data interoperability standards; and (iv) data retrieval and visualization solutions. The use of Blockchain technology in an EHR lifecycle is justified by its ability to guarantee the authenticity, traceability and security of data transactions processed by software systems. In this sense, Academia and the software industry have developed computational solutions using Blockchain technology that allow healthcare organizations to improve the quality of services provided to the population and reduce the operating costs involved in patient care. In the review carried out herein, it becomes evident that many business models in the healthcare sector are being transformed, and the use of Blockchain technology has impeded fraudulent alterations in EHR data and guaranteed the traceability of service provision agreements between patients and healthcare organizations. Analyzing the areas of an EHR lifecycle that are the object of this study, the following was found. The areas of storage and interoperability reflect most of the research analyzed, and this is justified by the need for data exchange that takes place between the stakeholders involved in the healthcare ecosystem. Meanwhile, the areas of data visualization, data modeling and EHR standardization require further investigation and the development of computational solutions. Based on the evidence collected in this study, our indications for future work include: (i) the specification of a method to help healthcare organizations in standardizing EHR data using the openEHR standard, and (ii) the development of a computational service for generating Blockchain transactions across different networks, and database management systems with Blockchain technology features.
References 1. P. Li et al., ChainSDI: A software-defined infrastructure for regulation-compliant home-based healthcare services secured by blockchains. IEEE Syst. J. 14(2), 2042–2053 (2020). https:// doi.org/10.1109/JSYST.2019.2937930 2. K. Kaur, R. Rani, Managing data in healthcare information systems: Many models, one solution. Computer 48(3), 52–59 (2015). https://doi.org/10.1109/MC.2015.77 3. A. Araújo, V. Times, M. Silva, A cloud service for graphical user interfaces generation and electronic health record storage, in Information Technology – New Generations, ed. by S. Latifi, (Springer, Cham, 2018), pp. 257–263
4. A. Araujo, V. Times, M. Silva, A tool for generating health applications using archetypes. IEEE Softw. 37(1), 60–67 (2020) 5. A. Araújo, V. Times, M. Urbano, PolyEHR: A framework for polyglot persistence of the electronic health record, in Int’l Conference Internet Computing and Internet of Things – ICOMP’16, pp. 71– 77,07 2016 6. M. del Carmen Legaz-García, C. Martínez-Costa, M. MenárguezTortosa, J.T. Fernández-Breis, A semantic web based framework for the interoperability and exploitation of clinical models and ehr data. Knowl Based Syst. 105, 175–189 (2016) 7. A. Clim, R.D. Zota, R. Constantinescu, Data exchanges based on blockchain in m-health applications. Proc. Comput. Sci. 160, 281– 288 (2019) 8. T.F. Stafford, H. Treiblmaier, Characteristics of a blockchain ecosystem for secure and shareable electronic medical records. IEEE Trans. Eng. Manage. 67(4), 1340–1362 (2020) 9. I. Lima, A. Araujo, R. Souza, H. Couto, V. Times, A service-based software architecture for enabling the electronic health record storage using blockchain. Int. J. Comput. Appl. Technol. 65(3), 222–234 (2021) 10. B. Shen, J. Guo, Y. Yang, Medchain: Efficient healthcare data sharing via blockchain. Appl. Sci. 9(6) (2019) 11. J. Chanchaichujit, A. Tan, F. Meng, S. Eaimkhong, Blockchain technology in healthcare, in Healthcare 4.0: Next Generation Processes with the Latest Technologies, (Springer, Singapore, 2019), pp. 37–62 12. W. Gordon, C. Catalini, Blockchain technology for healthcare: Facilitating the transition to patient-driven interoperability. Comput. Struct. Biotechnol. J. 16, 224–230 (2018) 13. S. Nakamoto, Bitcoin: A peer-to-peer electronic cash system, Cryptography Mailing list at https://metzdowd.com, 03 2009 14. S.S. Hazari, Q.H. Mahmoud, A parallel proof of work to improve transaction speed and scalability in blockchain systems, in 2019 IEEE 9th Annual Computing and Communication Workshop and Conference (CCWC), 2019, pp. 0916–0921. https://doi.org/ 10.1109/CCWC.2019.8666535 15. W. Li, S. Andreina, J.-M. Bohli, G. Karame, Securing proofof-stake blockchain protocols, in Data Privacy Management, Cryptocurrencies and Blockchain Technology, ed. by J. GarciaAlfaro, G. Navarro-Arribas, H. Hartenstein, J. Herrera-Joancomart, (Springer, Cham, 2017), pp. 297–315 16. K. Karantias, A. Kiayias, D. Zindros, Proof-of-burn, in Financial Cryptography and Data Security, ed. by J. Bonneau, N. Heninger, (Springer, Cham, 2020), pp. 523–540 17. A. Roehrs, C.A. da Costa, R.R. Righi, A.H. Mayer, V.F. da Silva, J.R. Goldim, D.C. Schmidt, Integrating multiple blockchains to support distributed personal health records. Health Inform. J. (2021). https://doi.org/10.1177/14604582211007546 18. P. Pandey, R. Litoriya, Securing and authenticating healthcare records through blockchain technology. Cryptologia 44(4), 341– 356 (2020) 19. P. Zhang, J. White, D. Schmidt, G. Lenz, S. Rosenbloom, Fhirchain: Applying blockchain to securely and scalably share clinical data. Comput. Struct. Biotechnol. J., 16, 267–278 (2018) 20. A. Dubovitskaya, Z. Xu, S. Ryu, M. Schumacher, F. Wang, Secure and trustable electronic medical records sharing using blockchain. AMIA Annu. Symp. Proc. AMIA Symp. 2017, 650–659 (2017) 21. H. Huang, X. Sun, F. Xiao, P. Zhu, W. Wang, Blockchain-based ehealth system for auditable ehrs manipulation in cloud environments. J. Parallel Distrib. Comput. 148, 46–57 (2021) 22. H. Guo, W. Li, M. Nejad, C. Shen, Access control for electronic health records with hybrid blockchain-edge architecture. IEEE Int. Conf. Blockchain (Blockchain) 2019, 44–51 (2019). https:// doi.org/10.1109/Blockchain.2019.00015 23. Y. Wang, M. He, CPDS: A cross-blockchain based privacypreserving data sharing for electronic health records, in 2021 IEEE
22 The Use of Blockchain Technology in Electronic Health Record Management: An Analysis of State of the Art…
24.
25.
26.
27.
28.
29.
30.
6th International Conference on Cloud Computing and Big Data Analytics (ICCCBDA), 2021, pp. 90–99, https://doi.org/10.1109/ ICCCBDA51879.2021.9442539 N. Islam, Y. Faheem, I.U. Din, M. Talha, M. Guizani, M. Khalil, A blockchain-based fog computing framework for activity recognition as an application to e-Healthcare services. Future Gener. Comput. Syst. 100, 569–578., ISSN 0167-739X (2019). https:// doi.org/10.1016/j.future.2019.05.059 G. Nagasubramanian, R. K. Sakthivel, R. Patan, A. H. Gandomi, M. Sankayya, and B. Balusamy, “Securing e-health records using keyless signature infrastructure blockchain technology in the cloud,” Neural Comput. Appl., vol. 32, pp. 639–647, Feb 2020. L. Chen, W.-K. Lee, C.-C. Chang, K.-K.R. Choo, N. Zhang, Blockchain based searchable encryption for electronic health record sharing. Future Gener. Comput. Syst. 95, 420–429 (2019) openEHR: Open industry specifications, models and software for e-health, https://www.openehr.org, 2021 [Online]. Accessed 12 Oct 2021 D. Bender, K. Sartipi, HL7 FHIR: An Agile and RESTful approach to healthcare information exchange, in Proceedings of the 26th IEEE International Symposium on Computer-Based Medical Systems, 2013, pp. 326–331, https://doi.org/10.1109/ CBMS.2013.6627810 A. Roehrs, C.A. da Costa, R.R. da Rosa, V.F. da Silva, J.R. Goldim, D.C. Schmidt, Analyzing the performance of a blockchain-based personal health record implementation. J Biomed Inform. 92, 103140 (2019). https://doi.org/10.1016/j.jbi.2019.103140. Epub 2019 Mar 4 A. Mohammad, R. Abdeen, T. Ali, Y. Khan, M.C.E. Yagoub, Fusing identity management, HL7 and Blockchain into a global healthcare record sharing architecture. Int. J. Adv. Comput. Sci. Appl. (IJACSA) 10(6) (2019). https://doi.org/10.14569/ IJACSA.2019.0100681
185
31. B. Alamri, I.T. Javed, T. Margaria, A GDPR-compliant framework for IoT-based personal health records using blockchain, in 2021 11th IFIP International Conference on New Technologies, Mobility and Security (NTMS), 2021, pp. 1–5, https://doi.org/10.1109/ NTMS49979.2021.9432661 32. BurstIQ – Whitepaper. https://www.burstiq.com/wp-content/ uploads/2017/09/BurstIQ-whitepaper_07Sep2017.pdf. 2017 [Online]. Accessed 12 Oct 2021 33. HealthNet – HealthNetTM : A Blockchain Enabled Population Health Platform for Enterprise Health Systems. https:// consensusnetworks.com/healthnet-whitepaper/ [Online]. Accessed 12 Oct 2021 34. MediLedger – Mediledger 2019 Progress report. https:// assets.chronicled.com/hubfs/MediLedger%202019%20Progress %20Report%20(1).pdf. 2019 [Online]. Accessed 12 Oct 2021 35. Corel Health – Coral Health The Blockchain for Personalized Medicine. https://ww1.prweb.com/prfiles/2018/07/31/15707263/ CoralHealth_WP.pdf [Online]. Accessed 12 Oct 2021 36. MedicalChain – Whitepaper 2.1 – https://medicalchain.com/ Medicalchain-Whitepaper-EN.pdf, 2018 [Online]. Accessed 12 Oct 2021 37. Nebula Genomics – Accelerating Genomic Data Generation and Facilitating Genomic Data Access Using Decentralization, Privacy-Preserving Technologies and Equitable Compensation. https://nebula.org/blog/wp-content/uploads/2019/05/AcceleratingGenomic-Data-Generation-and-Facilitating-Genomic-DataAccess.pdf. 2018 [Online]. Accessed 12 Oct 2021 38. A.M. Saghiri, Blockchain architecture, in Advanced Applications of Blockchain Technology, (Springer, 2019), pp. 161–176 39. A. Rubio, G.M. Tarazona, L. Contreras, Big Data and Blockchain Basis for Operating a New Archetype of Supply Chain (Springer, Cham, 2018), pp. 659–669
Blockchain for Security and Privacy of Healthcare Systems: A Protocol for Systematic Literature Review
23
Saadia Azemour, Meryeme Ayache, Hanane El Bakkali, and Amjad Gawanmeh
Abstract
Patient’s privacy, electronic health records’ confidentiality, integrity and all related e-health security issues are the most critical elements of a successful digital health, as they support building trust between patients and healthcare stakeholders. Blockchain technology appears to cover a wide range of these elements. However, the use of this emergent technology in healthcare domain, still has many security and privacy challenges that need to be overcome. Recently, many research works are focusing on such issues leading to a growing literature. In the perspective to review this literature in a systematic way to deeply investigate the use of blockchain technology in healthcare for security enhancement and privacy protection, this paper proposes a protocol that could be used to conduct successfully this systematic literature review (SLR). The proposed protocol follows the PRISMA-P 2015 Guidelines. At a closer look, we indicate the use of snowballing search and automated search (on eight electronic data sources) to carry out the intended SLR, identify five pertinent research questions, and specify the related inclusion/exclusion criteria. All methods for selection process, data collection and data S. Azemour () · H. E. Bakkali Smart Systems Laboratory- Rabat IT Center Mohammed V University Rabat, Rabat, Morocco e-mail: [email protected]; [email protected] M. Ayache STRS Laboratory, RAISS Team, INPT, Rabat, Morocco e-mail: [email protected] A. Gawanmeh College of Engineering and IT, University of Dubai, Dubai, UAE e-mail: [email protected]
analysis that will be used in the intended SLR are described in this protocol. Keywords
Blockchain technology · Smart contract · Distributed ledger · Security · Privacy · Interoperability · Healthcare · e-health · Telemedicine · Systematic literature review protocol
23.1
Introduction
23.1.1 Rationale Numerous systematic literature reviews deal with the use of blockchain in healthcare. For instance, [1, 2] investigate blockchain applications and platforms, and highlight challenges for possible future directions. Furthermore, [3, 4] explore the application domains of blockchain in healthcare, clarify the interaction of various healthcare stakeholders with the blockchain technology, and reveal some blockchain challenges and limitations. Moreover, [5, 6] treat briefly the security enhancement, privacy protection, scalability, interoperability, and disclose potential challenges. Many other studies explore only the security improvement based on the blockchain technology but in different domains for instance cryptocurrency, food supply chain, smart cities and Iot [7– 10]. We have found one systematic mapping study that deals with the use of blockchain technology in healthcare to improve security and privacy. However, this study does not conduct security challenges related to using blockchain technology in healthcare. On the other hand, few reviews were also found in this subject, but not systematically conducted [11–13].
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_23
187
188
S. Azemour et al.
Our final goal is to conduct a systematic literature review (SLR) to investigate the security and privacy challenges on healthcare. Therefore, and as suggested in [14–18], we started our SLR process with the planning stage that emphasize on developing a protocol for the study [19], then we will further conduct our SLR based on this protocol. Our SLR protocol follows the PRISMA-P 2015 guidelines [20,21]. Hence, we included all the 17 items that construct the PRISMA-P checklist, this will provide clarity, transparency and future reproducibility of our systematic literature review. As indicated in [22], the use of PRISMA-P 2015 facilitates the reporting of a completed systematic literature review.
23.1.2 Objectives The main research goals of the intended SLR are: (a) to gather knowledge about the enhancement of security and patient’s privacy in healthcare environment using the blockchain technology, (b) to assess the negative and positive influence of the blockchain technology in healthcare systems’ security and patient’s privacy, and finally (c) to spot potential research gaps and open issues that will be presented as future directions. This goal is broad enough to allow us to slice it in several more specific research questions. Using Kitchenham et al recommendations [18, 22], We have identified the five following research questions: • What are the security and privacy issues of the current used healthcare systems? • How the blockchain could improve security and privacy in digital health? • What are the existing blockchain applications that propose improvements for preserving security and privacy in digital health? • What are the limitations and the challenges of such applications (legal, human and technical aspects)? • How can we overcome those different limitations and challenges? The proposed protocol for our intended SLR, will allow us to further answer these research questions in the completed systematic literature review, by appealing to empirical results collected from target literature databases.
conducted on eight databases (ACM, Scopus, Springer, IEEE Xplore, Science direct, Google scholar, dblp, and web of science) using the defined search engine composed of a list of keywords related to blockchain, security and e-health. The second step of our methodology is related to the selection process which is based on the inclusion/exclusion criteria and includes several sub-steps: (a) firstly, we will reject all studies published before 2019, then (b) we will remove all duplicates, (c) finally, we will perform a selection based on the title, abstract and keywords, and keep only relevant papers that have a direct relationship with the SLR’s selected keywords. In the third step we will perform a full screen to the studies obtained from the previous step. This screen, is also based on the inclusion/exclusion criteria, keep only relevant papers after reading the full text. To support the selection process, we will further apply quality metrics that specify the information to be obtained in order to evaluate each study and eliminate studies with low quality. The last two steps are the extraction of data and the analysis of these data.
23.2.1 Eligibility Criteria To identify and select only relevant primary studies that focus on the area of interest, we will use the inclusion and exclusion criteria shown in Table 23.1. A paper will be excluded when it fulfills at least one of the exclusion criteria. Any paper not excluded by the exclusion criteria and fulfills all the inclusion criteria will be included in the set of selected primary studies. The application of the inclusion/exclusion criteria will be performed by reviewing meta-data information, such as title, abstract, keywords, index terms, etc. When decision cannot be made based on the available meta-data, then, conclusion contents will be considered in order to take the final decision [18] of inclusion/exclusion. The use of the proposed inclusion and exclusion criteria will prevent the introduction of bias into our selection process, appraise validity of the included primary studies and facilitate future reproducibility of our systematic literature review.
23.2.2 Information Sources
23.2
Systematic Literature Method
As indicated in Fig. 23.1 the empirical results will be collected by following several steps. The first step is the search process which utilize two types of search strategies in order to cover all relevant studies: (a) the snowballing search (backward and forward search which use 76 relevant articles as an initial set); (b) the automated search which will be
To ensure more complete coverage of our topic that encompasses all relevant studies, we will use two complementary search strategies snowballing search and automated search: • We will conduct a backward and forward snowballing, following Wohlin guidelines [23]. Initial set for snowballing will consist of some papers obtained from google scholar and identified as included studies. For backward search,
23
Blockchain for Security and Privacy of Healthcare Systems: A Protocol for Systematic Literature Review
ACM
Science Direct
Scopus
Databases Springer
Google scholar
dblp
189
Snowballing search IEEE Xplore
Web of science
Backward search
Forward search
Blockchain Keywords
Security Keywords eHealth Keywords
Refine by Date Use Inclusion/Exclusion Criteria
Eliminate Duplicates
Selecon process
Selecon based on tle, abstract, keywords
Evaluaon and quality assessment
Full text screening Data extracon Data analysis and Synthesizing finding
Fig. 23.1 The proposed methodology Table 23.1 Inclusion and exclusion criteria Inclusion criteria Papers published from 2019 to 2021 Papers written in English
Exclusion criteria Duplicated paper
Papers without full text availability Peer-reviewed papers published Short papers in an indexed conference proceeding or indexed journal Primary studies Secondary studies Papers related to the use of Papers that are technical reports Blockchain technology in or thesis healthcare for security and/or privacy purpose Papers that propose Blockchain frameworks or applications for enhancing security and/or privacy in healthcare
we will use the reference section of the examined work in order to identify papers that fulfill the inclusion-exclusion criteria and exclude others. For forward search, we will use google scholar to retrieve the citations and identify new papers to include (fulfill inclusion-exclusion criteria) from those papers citing the considered work. • The first step of the automated search is the selection of the Electronic Data Sources to be used [18, 22]. For us, we will use the following eight different electronic data sources: Google Scholar, IEEE Xplore, ACM Digital
Library, Scopus, SpringerLink, Science Direct, Web of science and dblp. The selected electronic data sources will cover almost every venue (Conferences, workshops and Journals) of software engineering field, which means that we can be confident that we will find almost all available evidences.
23.2.3 Search Strategy To obtain terms for the search string we have applied two different strategies: • firstly, We have analyzed our goal and research questions: As previously stated, our goal is to assess the use of blockchain technology in healthcare for enhancing security and protecting privacy. From this goal, we have extracted the terms: Blockchain, healthcare, security and privacy. These key terms also come from RQ1, RQ2, RQ3 respectively. In turn, RQ4 and RQ5 do not contain further key terms to add. • In the second step, We have conducted pilot searches: we have run pilot searches on google scholar, using the previous key terms, to identify other relevant terms that are frequently used. We looked for those terms in the title, Abstract and Author’s keywords of retrieved papers. For each key term we have found a set of relevant terms:
190
S. Azemour et al.
Fig. 23.2 Search string
Legend: Blockchain
AND Privacy
OR Healthcare
Security Distributed Ledger
e-health Cybersecurity
Smart Contract
MulChain
Aack Trust
m-health
Digital health
Vulnerability MulLedger
Telemedicine Interoperability
– Blockchain: Distributed Ledger, Smart Contract, Multi-chain, Multi-ledger. – Security: Privacy, Cybersecurity, Attack, Vulnerability, Trust, Interoperability – Healthcare: Healthcare, e-Health, m-Health, Digital health, Telemedicine. We connected the extracted terms using Boolean operators, likewise we used the operator AND to connect synonyms, and the operator OR to connect the sets, as detailed below: (Blockchain OR Distributed Ledger OR Smart Contract OR MultiChain OR Multi-ledger) AND (Security OR Privacy OR Cybersecurity OR Attack OR Vulnerability OR Trust OR interoperability) AND (Healthcare OR e-health OR m-health OR Digital health OR Telemedicine). For illustration purposes, Fig. 23.2 illustrates the list of terms used and their logical connection. Our search string was designed to find the maximum number of the known papers, However we may miss out some studies that use different terminology to describe the enhancement of security in healthcare based on blockchain. We have transformed the constructed search string into a search script adapted to each electronic data source as shown in the following example: • SpringerLink: (Blockchain OR “Distributed Ledger” OR “Smart Contract” OR MultiChain OR MultiLedger) AND (Security OR Privacy OR Cybersecurity OR Attack OR Vulnerability OR Trust OR interoperability) AND (Healthcare OR e-health OR m-health OR “Digital health” OR Telemedicine) We will use these scripts to perform our advanced search, based on specific metadata (title, abstract, keywords) in order to find only relevant papers that will be including in the
study. We will additionally specify the date range [2019– 2021] before starting the search process in each electronic data sources.
23.2.4 Study Record 23.2.4.1 Data Management Literature search results will be uploaded to Zotero [24] tool, an internet-based software program that provide bibliography management and results organization. We will create individual folders for each literature source on Zotero: backward-folder, forward-folder, and folder for each electronic data source (ACM-folder, Scopus-folder, etc.). Then we will create “exclusion-folder”, where we will place excluded articles, and we will specify the reason for exclusion of each article. Zotero will also help us to record the date of retrieval, abstract, keywords and notes about any other pertinent information. 23.2.4.2 Selection Process The same study may come from different sources, in consequence, we will firstly eliminate all duplicated studies, then we will perform an initial selection of the remainder studies based only on title in order to eliminate all unrelated papers, then we will review abstract to identify relevant papers that may provide direct evidence about the research questions, and fulfill the proposed inclusion criteria. However, sometimes the information given by these meta-data are mostly insufficient to determine whether a paper should be included or excluded. So that we will additionally read the conclusion for more information before making the final decision [18]. We attend to leave the full-text review of the included studies for quality assessment and data extraction process. The review’s selection process and full-text review will be illustrated using PRISMA 2020 flow diagram.
23
Blockchain for Security and Privacy of Healthcare Systems: A Protocol for Systematic Literature Review
As previously mentioned, we will record excluding papers and the reason to be excluded. The selection of all obtained studies will be conducted by One reviewer, then the selection results will be verified by two other reviewers, one of them will screen the included studies and the other one will verify the excluded studies, then a meeting will be performed to discuss the results and resolve any disagreements, this process will be confirmed by the supervisor in order to increase the reliability of our systematic literature review. Some authors could be contacted to resolve any remaining ambiguity.
23.2.5 Data Collection Process To ensure an efficient data extraction we will carefully follow the experience-based suggested guidelines [25]. We have additionally designed a data extraction form to accurately and systematically record the information obtained from included studies. As illustrated in Table 23.2, Data extracted will include general information (authors, title, publication type, country, etc), it will also contain study characteristics (methodology, objectives, contribution, etc) and information about the purpose of our SLR (current security issues, proposed blockchain solutions, security challenges of the use of blockchain in healthcare, etc). Table 23.2 Data extraction form Subtopics General information
Study characteristics
Data Title Authors Country Year of publication Publication type Publisher Citation count(from google scholar) Methodology
Study objectives Name of approach Contribution Project suitability Application domain Limitations Purpose of our Security issues in current healthcare systems SLR Blockchain solutions Blockchain Features related to security improvement Healthcare application domain Blockchain security challenges Corresponding stakeholders Security requirements Future directions
191
Data value, eligible for inclusion, research question and comments will be added to the data extraction form in order to assess the data extracted. Thus we will evaluate the eligibility to defined criteria (Yes, No or Unclear). if extracted data answers to any defined research question, it will be joined to this question. Comments will be added to record the text extracted from the study and contain this extracted data. Through the data collection process, we will read the full text of each included study, then extract relevant data and fill out the data extraction form that we have already designed, we will also add notes to each data extracted if necessary. For missing information, we will contact the included studies’ authors, by sending them three email attempts as a maximum. As with the selection process, single data extraction will be employed by one reviewer, then we will carry out a set of meetings with two other reviewers in order to discuss and verify the correctness of the obtained results and resolve disagreements. finally, our data collection will be confirmed by the supervisor in order to reduce bias and errors.
23.2.6 Data Items As previously stated, we will extract the main current security and privacy issues in healthcare system, and the blockchain security solutions that have been proposed to encounter these issues. Furthermore, we will collect blockchain security and privacy issues in healthcare and the most frequent attacks that may damage healthcare system or threat patient privacy. We will finally identify future directions. We intend to initially divide all the above-mentioned data items into five groups that present the healthcare stakeholders: patients, provider, research organization, supply chain bearer and payer. Each group will contain the corresponding current security issues, blockchain security solutions, blockchain security challenges and future directions. We will additionally categorize each study contribution as an application or framework or algorithm or any other proposal. All the blockchain security challenges will be identified based on the different security and privacy requirements.
23.2.7 Outcomes and Prioritization 23.2.7.1 Primary Outcomes The primary outcomes will be current security issues in healthcare, proposed blockchain security and privacy solutions, and finally blockchain security and privacy challenges. We decided a priori to subdivide security issues on current healthcare systems, blockchain solutions and blockchain security/privacy challenges into different healthcare stakeholders:
192
• Providers that may be hospital, doctor or specialist. • Patients. • Supply chain bearer that may be vendors, manufactures or pharmacy. • Research organization. • Payers (Insurance). Similarly, all those elements may be divided into different healthcare application domains, and may be joined to a specific security and privacy requirement, a potential scheme will be introduced to illustrate all these categorizations.
23.2.7.2 Secondary Outcomes The secondary outcomes will be attacks related to current security issues and blockchain security challenges, contribution type, the implementation or non-implementation of the proposed solution and research limitations. These additional outcomes will be reviewed to provide complementary information for the completeness of our systematic literature review. During this planning stage we will perform a sufficient investigation to ensure that the outcomes selected above are relevant. and we anticipate including these outcomes in a summary of finding table.
23.2.8 Risk of Bias Individual Studies We will use the Cochrane collaboration tool [26] to assess the possible risk of bias, we will additionally omit all studies that are judged to be high risk of bias. We will assess the possible risk of bias of each study by judging the extracted information as “high risk”, “moderate risk” or ‘low risk’. If there is insufficient detail reported in the study, we will judge the risk of bias as ‘unclear’ and the original study investigators will be contacted for more information. To facilitate the assessment of the overall risk of bias in included studies, we will use a checklist composed of all the predefined outcomes. We will omit all studies that are judged to be high risk of bias
23.2.9 Data Synthesis In this stage we will report studies with low risk of bias or moderate risk of bias, we also expect to retain studies at high risk of bias only when they provide a primary outcome. To answer to the previous research questions, a narrative qualitative summary will be performed at the first time [27]. Then, we will analyze the outcomes using sub group analysis to explore the possible source of heterogeneity based on the categorization indicated in data item and outcome subsections, we will additionally use tables and schemes to summarize and explain the finding characteristics. A comparison of the results will be performed. If collected data are not
S. Azemour et al.
in a suitable format for analysis, we will take various steps to process these data. We intend to perform a quantitative analysis, if included studies are sufficiently homogeneous in term of design. Thus, NVIVO software will be used for completing meta-analysis [28]. If possible, we will measure number of included studies per solution type, number of blockchain solutions per healthcare domain application, number of blockchain solutions per stakeholder, number of blockchain solutions per security requirement, and number of blockchain security challenges per healthcare domain application, and number of blockchain security challenges per stakeholder. When there are missing data, we will attempt to contact the original authors of the study to obtain the relevant missing data.
23.2.10 Confidence in Cumulative Estimate and Meta-Bias We will assess the quality of evidence for all outcomes using Grading GRADE (Grading of Recommendations Assessment Development and Evaluation), which is a system mostly used to judge the quality of evidence in healthcare literature [29]. The assessment will be performed across various domains of risk of bias, consistency, directness, precision and publication bias. we will additionally use an established approach that will assess the risk of bias across studies, and assess factors that increase the confidence in an effect based on the following quality questions: • Are the defined research objectives really match With our research questions? • How many research question are answered by the defined research objectives? • Is the contribution compared to other works? • Is the study contain future directions? We will judge the overall quality of evidence as high, moderate, low or very low. Evidences with low and very low quality will be omitted, and we will take only evidences with high and moderate quality.
23.3
Conclusion
In this paper we proposed a protocol that we will follow to successfully conduct our intended systematic literature review that will investigate the use of blockchain technology in healthcare for security enhancement and privacy protection. As a future work we will analyze the results of our systematic literature review conducted on the basis of this protocol. Thus, we will be able to identify the main healthcare security and privacy issues that blockchain technology
23
Blockchain for Security and Privacy of Healthcare Systems: A Protocol for Systematic Literature Review
has resolved, give various categorizations of the different blockchain solutions used for this purpose, and present the most important security and privacy challenges of the use of this emerging technology in healthcare. Therefore, we believe that our systematic literature review will make a valuable contribution to identify existing research gaps for future directions.
14. 15.
16.
References 1. M. Hölbl, M. Kompara, A. Kamišali´c, L.N. Zlatolas, A systematic review of the use of blockchain in healthcare. Symmetry 10(10), 470 (2018) 2. T.-T. Kuo, H.Z. Rojas, L. Ohno-Machado, Comparison of blockchain platforms: a systematic review and healthcare examples. J. Am. Med. Inf. Assoc. 26(5), 462–478 (2019) 3. Sobia, Y., M.M. Khan, R. Talib, A.D. Butt, S. Saleem, F. Arif, A. Nadeem, Use of blockchain in healthcare: a systematic literature review. Int. J. Adv. Comput. Sci. Appl. 10(5), 10 (2019) 4. L. Soltanisehat, R. Alizadeh, H. Hao, K.R. Choo, Technical, temporal, and spatial research challenges and opportunities in Blockchain-based healthcare: A systematic literature review. IEEE Trans. Eng. Manag. (2020). https://doi.org/10.1109/TEM.2020. 3013507 5. M. Prokofieva, S.J. Miah, Blockchain in healthcare. Aust. J. Inf. Syst. 23, 1–22 (2019) 6. A. Dubovitskaya, P. Novotny, Z. Xu, F. Wang, Applications of blockchain technology for data-sharing in oncology: 403–411. Results from a systematic literature review. Oncology 98(6), 403– 411 (2020) 7. A. Ekramifard, H. Amintoosi, A.H. Seno, A systematic literature review on blockchain-based solutions for iot security, in The 7th International Conference on Contemporary Issues in Data Science (Springer, Cham, 2019), pp. 311–321 8. Z. Yu, L. Song, L. Jiang, O. Khold Sharafi, Systematic literature review on the security challenges of blockchain in IoT-based smart cities. Kybernetes 51(1), 323–347 (2022) 9. A. Hassan, M.Z. Mas’ ud, W.M. Shah, S.F. Abdul-Latip, R. Ahmad, A. Ariffin, A systematic literature review on the security and privacy of the blockchain and cryptocurrency. OIC-CERT J. Cyber Secur. 2(1), 1–17 (2020) 10. N. Etemadi, Y.G. Borbon, F. Strozzi, Blockchain technology for cybersecurity applications in the food supply chain: a systematic literature review, in Proceedings of the XXIV Summer School “Francesco Turco” Industrial Systems Engineering, Bergamo, Italy (2020), pp. 9–11 11. A. Sarkar, T. Maitra, S. Neogy, Blockchain in healthcare system: security issues, attacks and challenges, in Blockchain Technology: Applications and Challenges (Springer, Cham, 2021), pp. 113–133 12. M. Soni, D.K. Singh, Blockchain-based security & privacy for biomedical and healthcare information exchange systems, Materials Today: Proceedings (2021) 13. S. Shi, D. He, L. Li, N. Kumar, M.K. Khan, K.-K. Raymond Choo, Applications of blockchain in ensuring the security andprivacy
17. 18.
19.
20.
21.
22. 23.
24. 25.
26.
27.
28. 29.
193
of electronic health record systems: a survey. Comput. Secur. 97, 101966 (2020) Y. Xiao, M. Watson, Guidance on conducting a systematic literature review. J. Plan. Educ. Res. 39(1), 93–112 (2019) P.V. Torres-Carrión, C.S. González-González, S. Aciar, G. Rodríguez-Morales, Methodology for systematic literature review applied to engineering and education, in 2018 IEEE Global Engineering Education Conference (EDUCON) (IEEE, Piscataway, 2018), pp. 1364–1373 J.L. Barros-Justo, S. Sepúlveda, N. Martínez-Araujo, A. GonzálezGarcía, The use of controlled vocabularies in requirements engineering activities: a protocol for a systematic literature review (2017). arXiv:1704.00822 A. Kofod-Petersen, How to do a structured literature review in computer science. Ver. 0.1 (2012) P. Brereton, B.A. Kitchenham, D. Budgen, M. Turner, M. Khalil, Lessons from applying the systematic literature review process within the software engineering domain. J. Syst. Softw. 80(4), 571– 583 (2007) C. Okoli, K. Schabram, A guide to conducting a systematic literature review of information systems research. Sprouts: Working papers on. Inf. Syst. 10(26) http://sprouts.aisnet.org/10-26 (2010) L. Shamseer, D. Moher, M. Clarke, D. Ghersi, A. Liberati, M. Petticrew, P. Shekelle, L.A. Stewart, Preferred reporting items for systematic review and meta-analysis protocols (PRISMA-P) 2015: elaboration and explanation. BMJ 349, g7647 (2015) D. Moher, L. Shamseer, M. Clarke, D. Ghersi, A. Liberati, M. Petticrew, P. Shekelle, L.A. Stewart, Preferred reporting items for systematic review and meta-analysis protocols (PRISMA-P) 2015 statement. Syst. Rev. 4(1), 1–9 (2015) B. Kitchenham, Procedures for Performing Systematic Reviews, vol. 33 (Keele University, Keele, 2004), pp. 1–26 C. Wohlin, Guidelines for snowballing in systematic literature studies and a replication in software engineering, in Proceedings of the 18th International Conference on Evaluation and Assessment in Software Engineering (2014), pp. 1–10 K.M. Ahmed, B. Al Dhubaib, Zotero: a bibliographic assistant to researcher. J. Pharm. Pharmacother. 4, 303–306 (2011) V. Garousi, M. Felderer, Experience-based guidelines for effective and efficient data extraction in systematic reviews in software engineering, in Proceedings of the 21st International Conference on Evaluation and Assessment in Software Engineering (2017), pp. 170–179 J.P.T. Higgins, D.G. Altman, P.C. Gøtzsche, P. Jüni, D. Moher, A.D. Oxman, , J. Savovi´c, K.F. Schulz, L. Weeks, J.A.C. Sterne, A.C. Jonathan, The cochrane collaboration’s tool for assessing risk of bias in randomised trials. BMJ J. 343, d5928 (2011) X. Huang, H. Zhang, X. Zhou, M.A. Babar, S. Yang, Synthesizing qualitative research in software engineering: a critical review, in Proceedings of the 40th International Conference on Software Engineering (2018), pp. 1207–1218 L.P. Wong, Data analysis in qualitative research: a brief guide to using NVivo. Malaysian Fam. Phys. 3(1), 14 (2008) G. Guyatt, A.D. Oxman, E.A. Akl, R. Kunz, G. Vist, J. Brozek, GRADE guidelines: introduction-GRADE evidence profiles and summary of findings tables. J. Clin. Epidemiol. 64(4), 383–394 (2011)
Single Sign-On (SSO) Fingerprint Authentication Using Blockchain
24
Abhijeet Thakurdesai, Marian Sorin Nistor, Doina Bein, Stefan Pickl, and Wolfgang Bein
Abstract
The objective of this paper is to describe the front-end and the backend of an open-source web-application that can be integrated in any website and for which the storage of single sign-on (SSO) authentication is provided in an Ethereum network. The backend of this application is shared with a browser-based platform or Android platform. Ethereum network facilitated the implementation of peer-to-peer multi-node blockchain in distributed ledger technology. We use smart contract code for user creation and authentication. A contract is a collection of code (its functions) and data (its state) that resides at a specific address on the Ethereum blockchain. The smart contract made our proposed web app self-verifying, self-executing, and tamper resistant. The proposed software system can be used as two factor authentications in combination with passwords for servers, for payments authorizations, in banking and automotive industry. Keywords
Android platform · Authentication system · Biometric data · Biometric trait · Distributed ledger technology · Ethereum network · Fingerprint data · Open-source A. Thakurdesai · D. Bein () Department of Computer Science, California State University, Fullerton, Fullerton, CA, USA e-mail: [email protected]; [email protected] M. S. Nistor · S. Pickl Department of Computer Science, Universität der Bundeswehr München, Neubiberg, Germany e-mail: [email protected]; [email protected] W. Bein Department of Computer Science, University of Nevada, Las Vegas, Las Vegas, NV, USA e-mail: [email protected]
web-application · Peer-to-peer multimode blockchain · Secure storage
24.1
Introduction
An authentication system provides the user a full-control and distributed access to authorized services, while protecting its the privacy of its identity. Generally speaking, authentication can be described as a process in which a user offers some form of proof that he is the same user who registered the account. When developing authentication systems for current web applications, the backend contains one centralized database in which the data used for authentication is being stored; for example using SQL, MongoDB, or MEAN stack. The user authenticates by providing the requested proof of identity; it is evaluated by the authentication server (part of backend) and depending on the result of this evaluation, the authentication server may or may not grant access to the user. A proof of identity can be any piece of information that an authentication server accepts: something users have in their possession, something they know or something they are (e.g., a password or a biometric) [1]. The oldest type of proof of identity is using passwords (strings of certain length), that can be stored as plain or encrypted. A breach disclosed by Equifax [2] has exposed identity information for over 140 million individuals. A biometric-based authentication system [3] requires first a user enrollment followed by, when needed, a user recognition. During user enrollment, a digital representation of an individual’s biometric trait is generated, called biometric template, then stored in a database for future comparison. The user recognition can operate in two modes: verification and identification. In the verification mode, the system require that the user provides a biometric template and this will be
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_24
195
196
matched against a single template in the database. In the identification mode, the system require that the user provides a biometric template and this will be matched against all templates stored in the database [1]. But even if the authentication system is biometric-based system, most of the deployed systems use a centralized model. in case an attacker gains access to the web application or the biometric centralized database, s/he can extract enough information to compromise the user’s digital identity [4]. If an authentication system uses a centralized architecture used for storage, then its data used for logging in and out of the system will be susceptible to attacks. And biometric data that can be used instead of textbased credentials is more valuable to attackers since it can provide access to many networks [1]. Since the user has a unique biometric trait, once extracted, it can be exploited to provide access to many networks and services. Single sign-on (SSO) enables the users to securely authorize and authenticate with multiple web or desktop applications using only a single pair of credentials. SSO schemes rely on a third-party identity provider (IdP) to broker authentication using protocols such as SAML [5] and OpenID Connect [6]. But only about 5% of sites use any of over 50 disparate IdP [7]. Loopholes in centralized IdP-based SSO systems are exploited and result in loss of biometric data [8]. If an application does not use a SSO then authentication, authorization and web application security has to be implemented and updated by the website itself including the user management part. SSO authentication works as follows (see Fig. 24.1): 1. A website checks if the user is already authenticated. If the user is already authenticated, then the content is delivered to the user. 2. If the user is not authenticated, then the website redirects the user to SSO login page. 3. The user enters the username and password to login. 4. The SSO identity servers verifies the credentials. 5. If the user credentials are correct, then the SSO server redirects the user to the website with appropriate HTTP code or else shows an invalid credentials error. The SSO in generally use a database to store user credentials which is secured by a web application. The SSO also has an admin dashboard for user management. So, the admin can add or remove access to the current users. Most of the SSO implementations available in the market currently are open source. The enterprise level SSO like OKTA includes 2-factor authentication in order to stop attacks on SSO system. But still the passwords make the systems prone to the hacking. Hence using an encrypted ledger like blockchain to store the data makes an excellent SSO. The blockchain is a type of distributed ledger, comprised of interchangeable, digitally recorded data in packages called blocks. Some advantages
A. Thakurdesai et al.
Fig. 24.1 SSO authentication flow
using blockchain for authentication include: it is impossible to tamper with the data, the central authority is eliminated, it is a decentralized ledger, and it has high concurrency to access data. In [1], the authors have proposed the Horcrux protocol as a method for secure exchange of biometric credentials within an existing standard (IEEE 2410-2017 BOPS) [9]. The proposed self-sovereign identity model uses blockchain technologies to assure exchange of verifiable credentials in lieu of 3rd-party identity providers during user authentication. The authors of [10] propose to secure an encrypted fingerprint template by a symmetric peer-to-peer network and symmetric encryption, Advanced Encryption Standard (AES) algorithm, and then is uploaded to a symmetrically distributed storage system, the InterPlanetary File system (IPFS). The encrypted template is uploaded to the IPFS, and its returned digest is stored on the Ethereum network. In this paper we will describe the front-end and the backend of an open-source web-application that can be integrated in any website and for which the storage is provided in an Ethereum network (see Fig. 24.2). We use fingerprint images as a biometric data to authenticate users. Fingerprints have been used for over a decade for user verification and identification [11–13]. The fingerprint collected during the
24 Single Sign-On (SSO) Fingerprint Authentication Using Blockchain
197
Fig. 24.2 Back-end modules for authentication
authentication needs to have a similarity score of at least 70% with the fingerprint images stored in the Ethereum network for the user to be successfully authenticated. If no object is having similarity score greater than 70, then a null object is returned in the form of custom message ‘http 404 User not found’ (see Fig. 24.4). Using the Ethereum network in our proposed system, data are stored on nodes that keep a copy of all transactions dating back to the first one in a chain of blocks. The hash of the previous block connects each block; hence, any tampering will be noticed. This system ensures data security when a node or multiple nodes are under attack. The major steps in developing our web app are: 1. Create a smart contract named ‘User.sol’ 2. Create and set up an Ethereum Network 3. Deploy the smart contract on the Ethereum network and store the address of smart contract 4. Create a web service using Java Springboot 5. Integrate the Web3j library to communicate between nodes of the Ethereum network 6. Install Postman to test the API 7. Access the smart contract using credentials and stored smart contract address. 8. Write a REST service named add_user to add a user 9. Write a REST service named verify_user to verify a user using fingerprint data. The entire project is stored as a Jar file. The paper is organized as follows. In Sect. 24.2 we present the backend, followed by the frontend in Sect. 24.3. Concluding remarks and future work are given in Sect. 24.4.
24.2
Backend of OUR Proposed Web App
The backend of this application is shared with a browserbased platform or Android platform. The core logic of the backend module is designed and developed using Springboot architecture, which a scalable way of creating the web
application. By using MVC architecture, we can decompose different modules of the application using independent service components. The application development involves designing and development of frontend, Restful web services to connect the front end and back end, and the backend module in Java. The MVC design pattern keeps the monolithic backend as modular as possible. This architecture can be further extended to microservices architecture where an application is built as small services, each of them being independently deployable. These services can be written in different programming languages and can use different databases for storage. The REST API initiates the transaction by communicating with the Java Tomcat REST server. Java Tomcat REST Server is responsible for Serving the Java web API. Web3j is intermediate component for communication between java RESTful API and Ethereum Network. Ethereum Network is the network of multiple blockchain nodes. We use smart contract code for user creation and authentication. A contract is a collection of code (its functions) and data (its state) that resides at a specific address on the Ethereum blockchain [14]. REST service add_user allows the admin to create a user. The admin can add users using a secure channel as follows. It first makes the request to Backend API to create a user. The backend API validates and serializes the biometric data, then it requests the Ethereum Network to store the data. An Ethereum node will mine the transaction and broadcast the resultant data to all the nodes in the Ethereum network (see Fig. 24.3). REST service verify_user authenticates the user and returns the corresponding object (aka biometric data) if matched. To authenticate a user, the following steps are executed. The user trying to authenticate will initiate a transaction and click on the “Verify” button (frontend), that will generate a verify request to the backend API. The backend API will extract the features from the input biometric object and request the Ethereum nodes to verify and provide some response. Ethereum nodes will provide details of all the objects with similar biometric features. The
198
A. Thakurdesai et al.
Fig. 24.3 Adding a user using an Ethereum network
backend API will then calculate the similarity score of all the returned objects and the single object having similarity score greater than 70 will be returned to the client. If multiple objects have similarity score greater than 70, then one with maximum score will be returned. If no object is having similarity score greater than 70, then null object is returned in the form of custom message ‘http 404 User not found’ (see Fig. 24.4).
authenticate has a similarity metric of 70% or higher, the backend will return a JSON object with the user’s details (see Fig. 24.9). If the details provided by the user do not manifest into the backend finding data, i.e. the user data is not present in the block chain database, the system returns a custom made error 404 error ‘User does not exist’ (see Fig. 24.10).
24.4 24.3
Frontend of the Proposed Web App
We implemented a very basic frontend, that can be expanded to any type of web app that is browser-based or Androidbased. Our basic frontend supports the backend functionalities, that include user creation and user authentication. Figure 24.5 shows the user registration webform to add a user. Once the fields are populated with necessary data, the user can be added (see Fig. 24.6). If the admin tries to add a user who already has biometrics registered, then the system throws an error “Operation not permitted” (see Fig. 24.7). This is done in order to avoid identity theft. The authentication of a user uses an API to verify its fingerprint (see Fig. 24.8). If a user trying to
Conclusion and Future Work
In this paper we describe the frontend and the backend of an open-source web-application that can be integrated in any website and for which the storage of single sign-on (SSO) authentication is provided in an Ethereum network. The backend of this application is shared with a browserbased platform or Android platform. An Ethereum network facilitated the implementation of peer-to-peer multi-node blockchain in distributed ledger technology. We use smart contract code for user creation and authentication. The smart contract code facilitated, verified, and enforced the negotiation and performance of an agreement and transaction. The smart contract made our proposed web app self-verifying, self-executing and tamper resistant. Our Web service makes the architecture to be RESTful and flexible.
24 Single Sign-On (SSO) Fingerprint Authentication Using Blockchain
199
Fig. 24.4 Biometric-based user authentication
Fig. 24.5 User registration screen
This projects demonstrates use of a distributed P2P structure in distributed computing, and demonstrate how secure a biometric-based authentication system can be made that it can replace passwords. The proposed software system can be used as two factor authentications in combination with passwords for servers, for payments authorizations, in banking and automotive industry. Two future directions are to implement microservices architecture for scaling and to implement an enterprise blockchain like IBM Hyperledger.
There are drawbacks to the proposed system that we plan to address in future work. One drawback is that, in order to keep the system running, the nodes of the system needs to be incentivized. The second drawback is that, if the private key of the seed node is lost then there is permanent loss of access. Acknowledgement This research was sponsored by the NATO Science for Peace and Security Programme under grant SPS MYP G5700.
200 Fig. 24.6 User registration is successful
Fig. 24.7 Denied registration of an existing user
A. Thakurdesai et al.
24 Single Sign-On (SSO) Fingerprint Authentication Using Blockchain Fig. 24.8 Webform to authenticate a user
Fig. 24.9 API for the user authentication is successful
201
202
A. Thakurdesai et al.
Fig. 24.10 API for user authentication does not recognize the user
References 1. A. Othman, J. Callahan, The Horcrux protocol: A method for decentralized biometric-based self-sovereign identity, in 2018 International Joint Conference on Neural Networks (IJCNN), Rio de Janeiro, 2018, pp. 1–7. https://doi.org/10.1109/ IJCNN.2018.8489316 2. M. Hume, Identity theft cited as threat after equifax security breach. Globe Mail Toronto A 7 (2004) 3. A. Jain, A. Ross, S. Prabhakar, An introduction to biometric recognition. IEEE Trans. Circ. Syst. Video Technol. 14(1), 4–20 (2004). https://doi.org/10.1109/TCSVT.2003.818349 4. A. Jain, K. Nandakumar, A. Nagar, Biometric template security. EURASIP J. Adv. Signal Proc., 2008(113), 1–17 (2008). https:// doi.org/10.1155/2008/579416 5. J. Hughes, E. Maler, Security assertion markup language (SAML) v2. 0 technical overview. OASIS SSTC Working Draft. sstc-samltechoverview-2.0-draft-08, 29–38 (2005) 6. N. Sakimura, J. Bradley, M. Jones, B. Medeiros, E. Jay. Openid connect standard 1.0-draft 18. Available online https://openid.net/ specs/openid-connect-standard-1_0-18.html. Last accessed 9 Dec 2020 7. A. Vapen, N. Carlsson, A. Mahanti, N. Shahmehri, A look at the third-party identity management landscape. IEEE Internet Comput. 20(2), 18–25 (2016)
8. K. Zetter, A. Greenberg, Why the opm breach is such a security and privacy debacle. Wired, 2015. Available online at https://www.wired.com/2015/06/opm-breach-security-privacydebacle/. Last accssed 9 Dec 2020 9. 2410-2017 IEEE biometric open protocol standard (BOPS). Available at https://standards.ieee.org/findstds/standard/24102017.html. Last accessed 9 Dec 2020 10. M.A. Acquah, N. Chen, J.-S. Pan, H.-M. Yang, B. Yan. Securing fingerprint template using blockchain and distributed storage system. Symmetry, pp. 1–13 (2020). Available online at https:// www.mdpi.com/2073-8994/12/6/951. Last accessed 9 Dec 2020 11. F. Hai-jian, C. Cheng-wei, Z. Chang-wei, Research on application of ethernet-based fingerprint identification system in college laboratory management, in Fourth International Symposium on Knowledge Acquisition and Modeling, Sanya, 1–274 (2011) 12. J. Tian, Y. Peng, Research of the Matlab application in the fingerprint identification system, in 2012 International Conference on Image Analysis and Signal Processing, Hangzhou, 2012, pp. 1–5. https://doi.org/10.1109/IASP.2012.6425005 13. S. Selvarani, S. Jebapriya, R.S. Mary, Automatic identification and detection of altered fingerprints, in 2014 International Conference on Intelligent Computing Applications, Coimbatore, 2014, pp. 239– 243. https://doi.org/10.1109/ICICA.2014.58 14. Ethereum, Introduction to Smart Contracts. Available at https://docs.soliditylang.org/en/v0.4.24/introduction-to-smartcontracts.html. Last accessed 9 Dec 2020
Part V Health Informatics
A Detection Method for Early-Stage Colorectal Cancer Using Dual-Tree Complex Wavelet Packet Transform
25
Daigo Takano and Teruya Minamoto
Abstract
Colorectal cancer is a major cause of death. As a result, cancer detection using supervised learning methods from endoscopic images is an active research area. Regarding early-stage colorectal cancer, preparing a significant number of labeled endoscopic images is impractical. We devise a technique for detecting early-stage colorectal cancer in this study. This technique consists of a 2D complex discrete wavelet packet transform and principal component analysis. As this technique does not require supervised learning, detection is feasible even in the absence of labeled data. In the endoscopic image, this technique correctly classifies early-stage colorectal cancer and normal regions with 92% accuracy. This approach outperforms the local binary pattern method. Keywords
Wavelet analysis · Dual-tree complex wavelet transform · Complex discrete wavelet transform · Complex discrete wavelet packet transform · Directional selectivity · Principal component analysis · Image classification · Early colorectal cancer · Endoscopic image
25.1
Introduction
If detected early, the five-year survival rate of colorectal cancer is exorbitant [1]. However, its detection is difficult. Endoscopy is the most common method of detecting colorectal cancer, and there are no definite detection criteria D. Takano () · T. Minamoto Graduate School of Science and Engineering, Saga University, Saga, Japan e-mail: [email protected]; [email protected]
for its early-stage by visual inspection. Early-stage colorectal cancer is diagnosed in regions different from the surrounding areas. Thus, the detection accuracy highly depends on the physician’s experience and knowledge. Applying supervised learning for cancer detection has recently become a routine [2], but detecting early-stage colorectal cancer is difficult. Generally, supervised learning requires a massive amount of training data and corresponding correct answer labels. Figure 25.1 shows that we should prepare an image with markings in the case of cancer 25.1, and data need to be generated by a physician. As mentioned earlier, detecting early-stage colorectal cancer is difficult, so creating data for supervised learning is challenging. We developed a detection method for early-stage colorectal cancer in this study; this method does not require supervised data. A 2D complex discrete wavelet packet transform (2DCWPT) and principal component analysis (PCA) were employed for feature extraction [3]. Further, to diagnose earlystage colorectal cancer, a thresholding process is applied to the obtained first principal component values. Early-stage colorectal cancer has a feature at the surface of the colon, which can be detected by observations from physicians. Therefore 2D-CWPT had directional selectivity can extract the difference in surface information between cancerous and normal regions. We assumed that the input is a single endoscope image in our detection method. For the given endoscopic image, we separated it into various regions and performed detection for each, and cropped 50 images of size 100 × 100 for each region. To get the high-frequency components, 2D-CWPT was used on the cropped images and the statistics from each high-frequency component.Next, we perform a PCA to these components. We described the 2D-CWPT used for feature extraction, and then proposed a method, which was followed by experimental results. Further, we applied our method to endoscopic images in the experiment. Using local binary pattern
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_25
205
206
D. Takano and T. Minamoto
filter was used in the row direction. L, H denotes the lowpass and high-pass components. LL means the low-frequency component, and LH, HL, HH means the high-frequency components in the vertical, horizontal, and diagonal directions, respectively. Equation (1) shows the decomposition of the scaling factor cRI j . H and G denote the low-pass and high-pass filters, respectively. cRI,LL j+1,ny ,nx =
hR2ny −ky hI2nx −kx cRI j,ky kx ,
ky ,kx
RI,LH dj+1,n = y ,nx
hR2ny −ky gI2nx −kx cRI j,ky kx ,
ky ,kx
RI,HL = dj+1,n y ,nx
(1) gR2ny −ky hI2nx −kx cRI j,ky kx ,
ky ,kx
Fig. 25.1 Example of an endoscopic image of early-stage colon cancer marked by a physician
[4], the classification accuracy of each cropped image was determined and compared with that of the existing method. This paper is organized as follows. First, we provide a brief description of 2D-CWPT in Sect. 25.2. We explain the proposed method and experiment for verification in Sects. 25.3 and 25.4. Finally, we describe the conclusion of the proposed method in Sect. 25.5.
RI,HH dj+1,n = y ,nx
2D Complex Discrete Wavelet Packet Transform
A method that applies wavelet packet to 2D complex discrete wavelet transform (2D-CDWT) is known as 2D-CWPT. However, as 2D-CWPT also decomposes high-frequency components, it is feasible to capture the variation of pixel values locally.
25.2.1 2D Complex Discrete Wavelet Transform 2D-CDWT is a wavelet transform that extends Kingsbury’s dual-tree complex wavelet transform (DT-CWT) to include full shift-invariance [5, 6]. It consists of two filter banks, one for each filter bank. Similar to the DT-CWT, it has two filters bank structures, one each for the real and imaginary parts. The filters applied in the row and column can be used to characterize the frequency components. Figure 25.2 shows the transform at level j = 1. The R and I represent the type of filter used in the column and row directions. Also C represents input image. For example, RI indicates that the real filter was used in the column direction and the imaginary
gR2ny −ky gI2nx −kx cRI j,ky kx
ky ,kx
By performing addition and subtraction, we can assign directional selectivity to a frequency component. Examples of adding direction selectivity to high-frequency component LH in the vertical direction is shown as follows: DR0,LH j,ny nx = DI0,LH j,ny nx =
25.2
DR1,LH j,ny nx = DI1,LH j,ny nx =
RR,LH II,LH + dj,n dj,n x ,ny x ,ny
2 IR,LH RI,LH dj,n − dj,n x ,ny x ,ny
2 RR,LH II,LH dj,n − dj,n x ,ny x ,ny
2
, , (2) ,
IR,LH RI,LH dj,n + dj,n x ,ny x ,ny
2
25.2.2 Dual-Tree Complex Wavelet Packet Transform 2D-CWPT is a wavelet packet technique that is applied to 2D-CDWT [3,7]. It decomposes high-frequency components during the conversion, unlike 2D-CDWT. The frequency components obtained from 2D-CWPT are characterized by the index (n, m). Here n represents the longitudinal frequency component ωy and m represents the transverse frequency component ωx . The higher the value of (n, m), the higher the frequency component. Figure 25.3 shows (n, m) for j = 1, 2. The conversion is conducted by 2D-CWPT according to the following rules: – If (n, m)=(1, 1), the real part of the filter is used. – If either of the indices n or m is an even number, low-pass and high-pass filters are switched.
25 A Detection Method for Early-Stage Colorectal Cancer Using Dual-Tree Complex Wavelet Packet Transform
207
Fig. 25.2 Flowchart of 2D-CDWT at level j = 1 to 2
be performed for each index (n, m). DR0,(n,m) = j,ny nx = DI0,(n,m) j,ny nx = DR1,(n,m) j,ny nx DI1,(n,m) j,ny nx
=
RR,(n,m) II,(n,m) + dj,n dj,n x ,ny x ,ny
2 IR,(n,m) RI,(n,m) dj,n − dj,n x ,ny x ,ny
2 RR,(n,m) II,(n,m) dj,n − dj,n x ,ny x ,ny
2
, , ,
IR,(n,m) RI,(n,m) dj,n + dj,n x ,ny x ,ny
2
Fig. 25.3 Index of each frequency component at levels 1 and 2 [3]
– If (n, m) = (1, 1), the filter used in 2D-CDWT is used. The transformation at (n, m) = (1, 2) with level j = 2 is shown in the following equation. The d( n, m)j in the equation represents the frequency component at level j. The notation rules for R and I are the same as for 2D-CDWT. RI,(2n−1,2m−1) = dj+1,n y ,nx
RI,(n,m) hR2nx −kx gR2ny −ky dj,k , y ,kx
25.3
Proposed Method
We explained the developed method for detecting earlystage colorectal cancer in this section and then described preprocessing performed on the endoscopic images. Further, preliminary experiments and their results were described. We used coefficient pn of the literature [5] in this study, and a range of 0 ≤ n ≤ 24 to create the filter.
ky ,kx
RI,(2n,2m−1) dj+1,n = y ,nx
RI,(n,m) hR2nx −kx hR2ny −ky dj,k , y ,kx
25.3.1 Preprocessing for Endoscopic Image
ky ,kx
RI,(2n−1,2m) dj+1,n = y ,nx
RI,(n,m) gR2nx −kx gR2ny −ky dj,k , y ,kx
ky ,kx
RI,(2n,2m) dj+1,n = y ,nx
RI,(n,m) gR2nx −kx hR2ny −ky dj,k y ,kx
ky ,kx
The same procedure used in 2D-CDWT can be used to add direction selectivity to 2D-CWPT. Note, this procedure must
We performed preprocessing to reduce the amount of strong light in the endoscope image, and surrounding regions on the left and right sides of the image folded back to the target area. Figure 25.4 shows a portion of Fig. 25.1 with this process applied. Using a single endoscope image, we cropped images for the proposed method. Figure 25.5 shows that we divided the
208
D. Takano and T. Minamoto
second principal component values plotted on each axis. Normal and abnormal regions are represented by blue and red dots, respectively, and a mixture of normal and abnormal region are represented by green dots. It can be observed that the first principal component values of abnormal regions data are greater than 0.
25.3.3 Proposed Detection Method Fig. 25.4 Image after preprocessing
We described our proposed method for detecting early-stage colorectal cancer in this section. We obtained an analytic image that applied the preprocessing of Sect. 25.3.1 to a single endoscopic image. Further, we applied 2D-CWPT and obtained the variance of each component with the exception of the low-frequency component. For each image, we created vectors with each variance as an element and used PCA. We calculated the probability of early-stage colorectal cancer by the percentage of the number of images with the first principal component value exceeding 0 for each region. We defined the first principal component as the degree of abnormality in this study.
Fig. 25.5 Block division For endoscopic image
25.4 endoscope image into several regions. Further, we cropped 50 analytical images with a size of 100 × 100 from each region. Each component image is converted from RGB to XYZ color space and graded to 15 grayscale as physicians perform endoscopy visually. The number of pixels, 100×100, and the Sturges formula was used to determine the gradation number. However, Y component images are omitted from the analysis because they indicate luminance. We applied the proposed method to images created using this method to detect cancer for each region.
25.3.2 Preliminary Experiment The ability of 2D-CWPT to capture the characteristics of early-stage colorectal cancer was confirmed, and the variance of each frequency component, calculated. From Sect. 25.2.2, in the transfor2D-CWPT generates a total of 4j+1 of D(n,m) j mation of level j. Among them, the frequency components belonging to (n, m) = (1, 1) are low-frequency components. In this study, we assumed j = 2 and calculated the variance of the 43 − 4 = 60 components, excluding the low-frequency components. Further, as we can obtain 60 variances from one endoscope image, the total number of data is N × 60 when the number of images is N. We applied PCA to the created data. Figure 25.6 shows a scatter graph of the endoscopic images, with the first and
Experimental Results
The outcomes of applying the proposed method are shown in this section. We tested the suggested approach on endoscopic images of colorectal cancer in its early stages, confirming that it can detect cancer. The detection accuracy of the suggested method was then compared to that of the local binary pattern [4].
25.4.1 Classification Experiment We applied the proposed method for Fig. 25.5. Figure 25.7 shows the probabilities calculated for each region. The region surrounded by a red frame in the figure indicates early-stage colorectal cancer, and the region surrounded by a blue frame is the normal region. However, the region surrounded a green frame indicates a mixed region by early-stage colorectal cancer and the normal region. The figure shows that earlystage colorectal cancer and the normal region can be detected correctly. One area was considered to be early-stage colorectal cancer with a probability of 24%, while the other was considered to be a normal area in the region where both areas were mixed. Next, we prepared 30 endoscopic images and confirmed the average of probability of early-stage colorectal cancer by cancer region, normal region, and mixed region. Table 25.1 shows this result. The average value was higher in the cancer region than in the normal region.
25 A Detection Method for Early-Stage Colorectal Cancer Using Dual-Tree Complex Wavelet Packet Transform Fig. 25.6 Examples of 1st to 2rd principal component values
209
Principle Component value
8
6
2nd principle component
4
2
0
–2
–4
–6
–8 –40
–20
0
20
40
60
80
100
120
1st principle component
Table 25.1 Probability of early-stage colorectal cancer for each region
Average of probability
Normal Region 0.22
Cancer Mixed Region Region 0.60 0.42
25.4.2 Comparison Experiment
Fig. 25.7 Anomaly rate of Image in Fig. 25.5
Further, we also confirmed which frequency components affect the first principal component. The correlation coefficient between each frequency component’s variance and the value of the first principal component was calculated. Figure 25.8 shows the frequency components whose correlation coefficient exceeded 0.9. This figure corresponds to the indexing in Fig. 25.3. In the II component, it is obvious that there are multiple frequency components with strong correlation coefficients. In particular, the high-frequency components in the diagonal direction of the II components all have high correlation coefficients. Therefore, we concluded that the characteristics of early-stage colorectal cancer exist there.
The method’s accuracy was compared with that of the existing method [4]. The existing method extracts features using discrete wavelet transform and local binary pattern and then applies support vector machine for detection. Further, we evaluated the proposed method’s accuracy with that of 2DCDWT rather than 2D-CWPT. Each method was applied after cutting 800 images from each of the two endoscope images. We confirmed the classification accuracy of each image for each method. Table 25.2 shows the accuracy of each method. We calculate accuracy(Acc), true positive(TP), and false positive(FP) for each method. Except for the TP in image 1, the suggested method outperforms the previous methods. Considering that the method [4] uses SVM for classification and has more parameters than the proposed method, it has a better detection method. We can also confirm that the proposed method outperforms the 2D-CDWT method in terms of accuracy.
210
D. Takano and T. Minamoto
Fig. 25.8 Frequency components with strong correlation
Table 25.2 Accuracy comparison between the proposed method and existing methods Method [4] Acca TPb Image1 77.8 76.5 Image2 87.2 79.2
2D-CDWT+PCA FPc Acc TP FP 79.0 77.8 55.0 100 94.0 92.5 83.6 100
Proposed Method Acc TP FP 78.0 56.0 100 92.7 84.0 100
a Accuracy b True c False
positive positive
25.5
Conclusion
In this study, developed a method for detecting early-stage colorectal cancer using 2D-CWPT and PCA. Our method does not require labeled data because it does not involve supervised learning. However, detection is possible even when data is insufficient. Further, we have confirmed that the high-frequency components in the diagonal direction of II components have features of early-stage colorectal cancer. Therefore, we can explain the reason for detecting by the proposed method. We compared the proposed method’s detection accuracy to that of the existing method [4] and confirmed that the proposed method’s accuracy was greater. Further, the proposed method’s accuracy was higher than that of a method that switched from 2D-CWPT to 2D-CDWT, confirming that 2D-CWPT was preferable for feature extraction in earlystage colorectal cancer. However, it is insufficiently accurate in regions that include both early-stage colorectal cancer and normal regions. Further, we have confirmed that the method is ineffective for images that do not contain cancer. Thus, we must find a more significant features of early-stage colorectal
cancer and develop an effective detection method for all endoscopic images. Acknowledgments This work was partially supported by JSPS KAKENHI Grant Number 19K03623.
References 1. American Cancer Society, 5-year relative survival rates for colon cancer (2021). https://www.cancer.org/cancer/colon-rectal-cancer/ detection-diagnosis-staging/survival-rates.html. Accessed Jan 29, 2021 2. Y. Sakai, S. Takemoto, K. Hori, M. Nishimura, H. Ikematsu, T. Yano, H. Yokota, Automatic detection of early gastric cancer in endoscopic images using a transferring convolutional neural network, in Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS, 2018-July, art. no. 8513274 (2018), pp. 4138–4141. https://doi.org/10.1109/ EMBC.2018.8513274 3. T. Kato, Z. Zhang, H. Toda, T. Imamura, T. Miyake, A novel design method for directional selection based on 2-dimensional complex wavelet packet transform. Int. J. Wavelets Multiresolution Inf. Process. 11(4), 1360010 (2013). https://doi.org/10.1142/ S0219691313600102 4. S. Charfi, M.E. Ansari, Computer-aided diagnosis system for colon abnormalities detection in wireless capsule endoscopy images. Multimedia Tools Appl. 77(3), 4047–4064 (2018). https://doi.org/ 10.1007/s11042-017-4555-7 5. H. Toda, Z. Zhang, Perfect translation invariance with a wide range of shapes of Hilbert transform pairs of wavelet bases. Int. J. Wavelets Multiresolution Inf. Process. 8(4), 501–520 (2010). https://doi.org/10.1142/S0219691310003602 6. R.R. Coifman, Y. Meyer, V. Wickerhauser, Wavelet analysis and signal processing, in Wavelets and Their Applications (Springer, Berlin, 1992) 7. N. Kingsbury, Complex wavelets for shift invariant analysis and filtering of signals. Appl. Comput. Harmonic Anal. 10(3), 234–253 (2001). https://doi.org/10.1006/acha.2000.0343
Visualizing 3D Human Organs for Medical Training
26
Joshua Chen, Paul J. Cuaresma, Jennifer S. Chen, Fangyang Shen, and Yun Tian
Abstract
Three-dimensional (3D) models have been used as essential tools in medical training. In this study, we visualize 3D models of human organs with graphics software for the purpose of training medical students. This study investigates whether 3D organ visualizations will be more recognizable to medical students than two-dimensional (2D) organ images. In our experiments, the models were shown to health science students to determine how useful they were in training and we compared the use of 3D models with 2D images. We conclude that the 3D organ models we used are more likely to be recognized by the students. Keywords
3D · Visualization · Medical training · Blender · 3D human organs · 2D organ images · Medical education · STL files · Health Science · 3D models J. Chen Mililani High School, Central School District of Mililani Town, Mililani, MI, USA P. J. Cuaresma Guam High School, DoDEA Pacific West District, Agana, Guam e-mail: [email protected] J. S. Chen Medical Education International, Christian Medical Dental Association, Mililani, MI, USA F. Shen Department of computer system tech., New York City College of Technology, New York, NY, USA e-mail: [email protected] Y. Tian () Department of Computer Sciences & Electrical Engineering, Eastern Washington University, Spokane, WA, USA e-mail: [email protected]
26.1
Introduction
Three dimensional (3D) visualization provides a way for users to interactively examine a 3D object. Technologies such as volume renderings, are adopted to visualize CT scans as a 3D model, providing physicians with more data to determine the course of treatment for patients [1, 2]. 3D models provide surgeons with more detailed information, thus enhances confidence in their decisions when operating on the patient. For example, a patient may have a crushed pelvis from a car accident. After being able to see the patient’s pelvis from all angles using a 3D model, the doctors are able to resolve the best way to operate on the patient. Simulations based on 3D models play a critical role in the medical field. Simulation may serve as a training technique that replaces real experiences with synthesized ones [3–5]. It allows people to practice repeatedly on a computer until they have mastered the skills, which gives the students hands-on experience. In other situations, simulation can help medical students develop skills that would otherwise entail quite scarce materials or opportunities [6]. For example, simulation can be used as an alternative to using real cadavers and live patients. This is important because cadavers are scarce resources [7]. Sometimes, real patients are used to help trainees practice medical techniques. As a result, those patients may have to receive invasive operations [8]. Medical students also use simulations to practice surgical procedures in rare, unusual, or high risk situations that live patients cannot offer [9, 10]. The widely applied 3D objects and models in health science motivate us to adopt 3D visualizations techniques in fundamental training for medical students. In this work, we attempt to verify whether health science students with a basic understanding of human anatomy will show higher accuracy
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_26
211
212
J. Chen et al.
and confidence when attempting to identify 3D models of human organs as opposed to two-dimensional (2D) images. In this study, we visualize 3D models of human organs with graphics software for the purpose of training medical students. This study investigates whether 3D organ visualizations will be more recognizable to medical students than 2D organ images. In our experiments, the models were shown to health science students to determine how useful they were in training and we compared the use of 3D models with 2D images. We conclude that a majority of 3D organ models we used are more likely to be recognized by the students. The rest of the paper is organized as follows. Section 26.2 presents our methodology and the data set in our research. In Sect. 26.3, we present and discuss our experiment results. We conclude by discussing the limitations, our conclusions and future work in Sect. 26.4
26.2
Fig. 26.1 3D model of a human heart
Research Methodology
26.2.1 Overview In the project, 3D human organ models were visualized with Blender, a free and open-source 3D visualization software that can be used for modeling, rendering, and animation [11– 13]. 3D models of human organs in the format of .stl files were downloaded from embodi3.com, sketchfab.com and turbosquid.com. Six 3D medical models were employed in the experiments. We created several short videos by screenrecording the rotations and translations of the 3D models when interacting with them in Blender. Then, we present to two classes of health science students the videos of the 3D models, together with some 2D images of the same set of human organs. In a survey, multiple choice questions were given to the health science students, asking what human organs they have observed in both the videos that contain the 3D models and in the images of 2D human organs. We reported the accuracy of recognizing the human organs and the confidence level of the students.
26.2.2 Experiment Setup By default, the 3D models that we downloaded from the websites were not rendered on the screen with realistic color and lighting settings, as shown in Fig. 26.1. We improved the 3D models with more realistic colors and lighting parameters in the visualization using Blender. A picture of the improved 3D model is shown in Fig. 26.2. Blender software was installed and used on macOS Catalina with a version 10.15.2. The videos of visualizing the 3D models along with the survey questions were posted at the web URL: https://create.kahoot.it/details/
Fig. 26.2 The improved 3D model of the human heart
group-2-a/658bdc38-06c0-40dc-9560-94a5b8964066. The corresponding 2D images of the same set of human organs along with survey questions were posted at https:// create.kahoot.it/share/group-1-a/54c345ce-a042-4d3eb16c-3f25fa3eec88.
26.2.3 Experiment Implementation Two sets of surveys were given two different classes, Health Science I and Health Science II. There were 23 students in Health Science I and 6 students in Health Science II. Feedback from each class was collected and analyzed. The specifics are described below. First, a six-second long video was recorded when a model were rotating during visualization to show all sides of the model. These videos
26 Visualizing 3D Human Organs for Medical Training
were uploaded to YouTube so they could be inserted into a Kahoot quiz. Second, two Kahoot quizzes/surveys were created, referred to as 3D-model survey and 2D-image survey respectively. Third, the students were informed about the survey and were given instructions before taking the surveys. Specifically, within each class, students were divided into two groups, group One and group Two. Group One was given the 3D-model survey, while group Two was given the 2Dimage survey. This was done to minimize the effects between the students on who would be shown the 3D models first or vice versa. When the students in group One were taking the 3D-model survey, the students in group Two turned around to face away from the screen. Then, the students in group Two took the 2D-image survey, when students in group One faced away from the screen. During the survey time, the videos and images were shown on a large screen and the students answered the questions in the predesigned survey. Each survey question were displayed for five seconds and ten seconds were given to the students to choose what they thought a 3D model or a 2D image was represented between four multiple choice answers. Then, the survey asked the students to rate their confidence level on a four-point scale within a twenty-second limit. This was repeated for every model. Two days after the first survey were given, group One was given the 2D-image survey and group Two was given the 3Dmodel survey. The students had to wait two days before they took the other survey, because they needed enough time to forget what they had already seen. The students were also instructed not to share the answers to the survey questions. Seventh, feedback from each class was collected from the kahoot.com website and analyzed. In the experiment, the independent variables were the 3D models of human organs. The color and shape of the models can be changed to make it easier for the health science students to identify the models. The dependent variables are the accuracy of the students’ answers and their confidence level. The students were a random variable because every student was different and could not be controlled. Each student had a distinct level of knowledge of the human body and a different level of confidence. The constants in this experiment were the survey that the students took and the class the students were enrolled in. Each student had the same set of models/images, amount of time, and multiple choice answers to choose from. The results were separated by different classes because every class has a different curriculum and different interests. These differences have an impact on the students’ knowledge of human anatomy and would affect the results of the survey.
213
26.3
Experimental Results
We present the experimental results in this section, including the findings as to the adoption of the 3D models versus the 2D images.
26.3.1 Results for Our 3D Models The recognition correctness rate are shown in Figs. 26.3 and 26.5 for two different Health Science classes. The percentages were calculated by dividing the number of students who gave the correct answer by the total number of students. The corresponding confidence level of the student participants are presented in Figs. 26.4 and 26.6.
26.3.2 Analysis and Discussion This study attempts to verify whether 3D models will be more effective to be recognized in medical training. This was done by comparing how accurately health science students can recognize human organs represented in 2D images against 3D visualizations. As shown in Figs. 26.3 and 26.5, results indicate that the 3D models are very easily recognized because all but three of the tests of the 3D models had at least 90% of all answers correct. Moreover, the results from the Health Science I class showed that on average, the use of 3D models achieved a marginal higher correctness rate than that of using the 2D images. The correctness rate for the 3D models compared to 2D images increased by an average of 3%, as shown in Figs. 26.3 and 26.5. For the students in the Health Science I class, as shown in Fig. 26.3, the 3D models have been recognized with a higher accuracy than the 2D images, except for the Lungs and Kidney. However, as shown in Figs. 26.4 and 26.6, on average, the students who answered correctly indicated higher confidence when using the 2D image representations. The survey from the Health Science II class also showed higher accuracy rates when the 2D representations were presented. The pattern was opposite to what is expected. It was expected that the 3D models would be more accurate and have more confidence because they could be viewed from different angles. This could have occurred because the textbooks that were used in the class all show 2D representations of organs. The students
214 Fig. 26.3 Recognition Accuracy for the Health Science I Class
Fig. 26.4 Average Confidence Level of The Students in Health Science I
Fig. 26.5 Recognition Accuracy for the Health Science II Class
J. Chen et al.
26 Visualizing 3D Human Organs for Medical Training
215
Fig. 26.6 Average Confidence Level of The Students in Health Science II
would be more used to the 2D images and were able to recognize the 2D images more easily. In addition, the 3D kidney had less accuracy than all of the other 3D organs. This could be also because the kidney was shown to the students first and the students had to get used to viewing a 3D model. The experiment could be improved by showing some 3D models that were not used in the experiment for warm-up purposes, before showing the medical models in the experiment. Moreover, the more advanced class was able to answer the questions with almost 100% of accuracy for both 2D images and 3D models, as observed in the Health Science II class. This could be because the advanced students had more knowledge of what organs look like and did not see a difference between the 2D and 3D representations. Because they had more experience with identifying human organs, they were able to identify the images effortlessly. This experiment could be improved by making the test harder for advanced students in order to better see the difference between 2D images and 3D models. To summarize, on average, using the 3D representations had the same amount of or more correct answers than the 2D images. Some of the 3D visualizations are able to replace 2D images as learning tools while other 3D models still need to be improved.
26.4
Conclusions and Future Work
In this work, we visualize 3D models of human biological structures using graphics software. The results showed that the 3D models achieved the same amount of or more accurate recognition rate than that of the 2D images. In the future, the 3D models could be further improved by adding more details, specifically changing the color, shape, resolution, and
lighting to make the models more realistic. For example, parts of the 3D models may be made transparent and allow the viewers to see the inside of the organ. We can also consider using other software programs to visualize the 3D models.
A.1
Appendices
A.1.1 Appendix A: The 3D Models Used in the Survey
216
A.1.2 Appendix B: The 2D Images Used in the Survey
J. Chen et al.
26 Visualizing 3D Human Organs for Medical Training
217
References 1. G. Zörnack, J. Weiss, G. Schummers, U. Eck, N. Navab, Evaluating surface visualization methods in semi-transparent volume rendering in virtual reality. Comput. Methods Biomech. Biomed. Eng. Imaging Visual., 1–10 (2021) 2. Grant, Gerald T. “3D volume rendering and 3D printing (additive manufacturing), in Emerging Imaging Technologies in DentoMaxillofacial Region, An Issue of Dental Clinics of North America, E-Book 62(3), p. 393. (2018). 3. O. Chernikova, N. Heitzmann, M. Stadler, D. Holzberger, T. Seidel, F. Fischer, Simulation-based learning in higher education: a metaanalysis. Rev. Educ. Res. 90(4), 499–541 (2020) 4. J. Kim, C. Choi, S. De, M.A. Srinivasan, Virtual surgery simulation for medical training using multiresolution organ models. Int. J. Med. Robot. Comput. Assist. Surg 3(2), 149–158 (2007) 5. H.M. Nassar, A. Tekian, Computer simulation and virtual reality in undergraduate operative and restorative dental education: A critical review. J Dent Edu. 84(7), 812–829 (2020) 6. R. Umoren, S. Bucher, D.S. Hippe, B.N. Ezenwa, I.B. Fajolu, F.M. Okwako, J. Feltner, et al., eHBB: a randomisedcontrolled trial of
7.
8.
9. 10. 11. 12. 13.
virtual reality or video for neonatal resuscitation refresher training in healthcare workers in resource-scarce settings. BMJ open 11(8), e048506 (2021) H. Brenton, J. Hernandez, F. Bello, P. Strutton, S. Purkayastha, T. Firth, A. Darzi, Using multimedia and Web3D to enhance anatomy teaching. Comput. Educ. 49(1), 32–53 (2007) R.J. Scalese, V.T. Obeso, S. Barry Issenberg, Simulation technology for skills training and competency assessment in medical education. J Gen. Inter. Med. 23(1), 46–49 (2008) P.A. Guze, Using technology to meet the challenges of medical education. Trans. Am. Clin. Climatol. Assoc. 126, 260 (2015) R.A. Agha, A.J. Fowler, The role and validity of surgical simulation. Int Surg 100(2), 350–357 (2015) E.T. Guevarra, Mendoza., Modeling and animation using blender: Blender 2.80: The rise of eevee (Apress, 2019) M. Gárate, Voxel datacubes for 3D visualization in Blender. Public. Astron Soc Pacif 129(975), 058010 (2017) J.D. Durrant, BlendMol: Advanced macromolecular visualization in Blender. Bioinformatics35(13), 2323–2325 (2019)
An Information Management System for the COVID-19 Pandemic Using Blockchain
27
Marcelo Alexandre M. da Conceicao, Oswaldo S. C. Neto, Andre B. Baccarin, Luan H. S. Dantas, Joao P. S. Mendes, Vinicius P. Lippi, Gildarcio S. Gonçalves, Adilson M. Da Cunha, Luiz A. Vieira Dias, Johnny C. Marques, and Paulo M. Tasinaffo
Abstract
During the 1st Semester of 2020, 25 students from the Aeronautics Institute of Technology (Instituto Tecnológico de Aeronáutica – ITA) in São José dos Campos, SP, Brazil developed the academic project “Specific Technological Solutions for Special Patients with Big Data”, in Portuguese Projeto STEPES-BD: Soluções Tecnológicas Específicas para Pacientes Especiais e Sistemas em Bancos de Dados. They have accepted the challenge of using a technological approach to help manage and combat the Sars-CoV-2 Virus Pandemic (COVID-19). At that time, the lack of shared data between public and private agencies and the need for faster information flow were considered the main information for combating the spread of the disease to guide the development of an Information Management System for the COVID-19 Pandemic, involving the essential data from Electronic Health Records (EHRs). The combination of some emerging technologies like Big Data and Blockchain, together with the Scrum Framework (SF) and the Interdisciplinary ProblemBased Learning (IPBL) enabled those students from three different academic courses to develop a computer system prototype based on the pressing needs caused by this disease. This article describes the development of the main deliverables made by those graduate students in M. A. M. da Conceicao () · O. S. C. Neto · A. B. Baccarin L. H. S. Dantas · V. P. Lippi · G. S. Gonçalves · A. M. Da Cunha L. A. Vieira Dias · J. C. Marques · P. M. Tasinaffo Computer Science Division, Aeronautics Institute of Technology – ITA, São José dos Campos, Brazil e-mail: [email protected] J. P. S. Mendes State Technology Educational Center Paula Souza, São José dos Campos, Brazil
just 17 academic weeks right from the beginning of the COVID-19 Pandemic crisis in the 1st Semester of 2020. Keywords
Software Blockchain · Interdisciplinary education · Problem-based learning · Electronic health records · Scrum · Big Data · Database · Descentralized control · Digital assets · Public healthcare
27.1
Introduction
This article presents an experience of students from the Aeronautics Institute of Technology (Instituto Tecnológico de Aeronáutica – ITA) in the development of the academic project “Specific Technological Solutions for Special Patients with Big Data”, in Portuguese Projeto STEPES-BD: Soluções Tecnológicas Específicas para Pacientes Especiais e Sistemas em Bancos de Dados, applying Interdisciplinary Problem-Based Learning (IPBL) for monitoring patients with suspected COVID-19. The STEPES-BD project aimed to develop a computer system, applying knowledge from the development of a practical and interdisciplinary project as a Proof of Concept (PoC). It involved students from 3 different courses of the Graduate Program in Electronic and Computing Engineering in the Area of Informatics (Programa de Pós-Graduação em Engenharia Eletrônica e Computação na Area de Informática – PG/EEC-I), using the Scrum Framework (SF) in an academic environment together with some emerging technologies like Big Data, Blockchain, among others. This project was developed by 25 students from the following courses on the 1st Semester of 2020: CE-245 Information Technologies; CE-240 Database Systems Project; and CE-229 Software Testing. Due to the COVID-19 Pandemic,
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_27
219
220
M. A. M. da Conceicao et al.
the project was carried out remotedely through meetings by video conference. The project addressed the problem of the Sars-CoV-2 virus also known as COVID-19. To carry out efficient monitoring of infected patients, an Information Management System was developed capable of sharing data between the main actors of a Health System, aiming to provide different views on the spread of the disease and appropriate decision making. One of the impacting factors found in the process of combating the COVID-19 Pandemic was the ineffectiveness of sharing data from the main actors of the process in the Health System: the Government and the Society [1]. This sharing involves dependencies, compilations, periodic updates, and monitoring of real data, based on real-time data collections considered essential for decision making. The impacts of this virus in the supply chain and health services caused price increases and product shortages, highlighting the need for developing a system that would provide traceability of data to make up the logistics of inputs and health products, to improve the applicability of involved resources from the public and private sectors [2]. Within this context, the STEPES-BD project was specified to provide the monitoring of infected patients, through appropriate data sharing and medical records, using Electronic Health Records (EHR) [3], creating standards and providing real-time health data analysis. This Project was conceived and developed to make data related to the hospital supply chain more transparent, providing traceability and immutability of data in a real environment, using Big Data, Blockchain, and other emerging technologies. In addition, this prototype-level system design mainly involved: Scalability of Relational and Non-Relational Databases; Big Data; Cloud Computing; Internet of Things (IoT); as well as a real environment for applications of Software Testing methods and techniques.
27.2
Tailoring the Scrum
Due to the social distance required by the COVID-19 Pandemic, it was necessary to adjust the Totally Integrated Scrum (TIS) method to the development of the project by the students, according to the Scrum model for Distributed Agile Development [4]. In this adaptation to a TIS, the Scrum Teams (STs) were considered to be multi-functional and their members were geographically distributed, requiring frequent and periodic inter-team communications and adopting Scrum of Scrums meetings.
As it is an academic project with well-defined schedule restrictions in just 17 weeks and of variable scope, the Project management was carried out using the best practices of Scrum. For that, the time boxes, the intervals between the ceremonies, and the artifacts had to undergo adaptations, to allow students more opportunities for contacts with the agile software development management method. To meet these adaptations, the STEPES-BD project was divided into four segments, each of which being assigned to the following STs: Patients; Physicians; Hospitals; and Suppliers. Although the STs had specific scopes in their segments, they started to work together since the Sprint 2 in an interdisciplinary effort to address activities related to integrations, involving the different needs of infrastructure and technologies of Big Data, Blockchain, among others. The selection and distribution of the students by the STs were carried out with the participation of students and professors soughing to balance quantitative and qualitatively the members of each ST, adjusting their skills, qualifications, and previous experiences. In addition, the following Scrum Roles exercised by the students were also adapted, when two new roles were created and other two renamed within the project: – General Product Owner (GPO) – The Product Owner (PO) role was renamed as General Product Owner (GPO); – Backup of General Product Owner (Bkp of GPO) – This new academic role was created to replace the GPO and provide immediate answers, in case of absences of the General Product Owner (GPO); – Team Scrum Masters (TSMs) – The role of Scrum Master (SM) was renamed as TS Master (TSMs); and – Backup of Team Scrum Master (Bkp of TSM) – This new academic role was created to replace the TSM and to provide more immediate responses in the eventual absence of the TSMs. Professors acted as Stakeholders and monitored the development of the computer system by reviewing the artifacts generated by the students, periodically auditing: the Product Backlog; Sprints’ Backlogs; Sprints’ Kanbans; Test Cases’ Spreadsheets; among the minimum, necessary, and sufficient documents, by using Burndown Charts metrics. To assess the levels of maturity reached by the STs in the development of their Segments, involving the ceremonies performed and the preparation of the artifacts, the Tuckman Maturity Model (TMM) was used. Thus, the STs based on self-management began to be periodically checked and monitored for their evolution in maturity levels, as the project was being developed, allowing behavioral realignments of its members throughout the four stages of the TMM: Forming, Storming, Norming, and Performing [5].
27 An Information Management System for the COVID-19 Pandemic Using Blockchain
27.3
Background
This section describes the built architecture and highlights the main technologies and tools used in the development of the STEPES-BD project, in terms of its Proof of Concept (PoC).
27.3.1 The STEPES-BD Architecture The conceived architecture was based on the N-Tiers concept (development in layers) [6] and allowed the division of the Project into the Layers of: Interface; Business; and Persistence, as shown in Fig. 27.1. 1. Interface Layer: This layer was responsible for the interaction and presentation of the system to the user and had modules addressed to the functions of Registries, Medical Assistances, Hospital Supplies, and Dashboards. To provide data sharing between the STEPES-BD subsystems, the interoperability was implemented by using the RE presentational State Transfer (REST), a style of software architecture that defines a set of restrictions for the creation of Web Services in Application Programs Interface (APIs), as shown in Fig. 27.2. 2. Business Layer: In this layer, it was possible to achieve data interdependence for the completeness of the processes proposed by the STEPES-BD project, through flexibilities obtained from the open API routes in the subsystems, providing fast and efficient exchanges of information managed by the development teams. 3. Persistence Layer: In this layer, involving data transacted by domain entities, the MySQL Relational Database was used and the BigchainDB was introduced, as a database structured in Blockchain that made possible the use of this technology in data management and information recording, foreseeing controls through a ledger, for future conferences and guarantees of data traceability. The system proposed in the STEPES-BD project made extensive use of the data managed in the Blockchain blocks and the mechanisms for searching, consulting, and locating medical records and contracts for the purchase and sale of hospital supplies have used the tools provided by the BigchainDB API. The BigchainDB worked on the STEPES-BD project in a MongoDB database that, through its infrastructure, made it possible to store a framework that makes up the Blockchain blocks. For that, the BigchainDB has created
221
on MongoDB the collections of: Assets, Transactions, and Metadata. The BigchainDB also enabled a wide use of the concept of Assets, representing any item of value [7]. In this sense, the Collection Assets of the database kept the most valuable records inside. In this location, materials transacted in the hospital supply logistics chain were stored. From then on, all items could be properly identified by serial numbers, lots, or any other forms of identification; and by medical records generated in consultations, containing relevant data such as identification of those involved, diagnoses, hospital items consumed during consultations, among others. Subsequently, the Assets were encapsulated within structures in the JavaScript Object Notation (JSON) format to be traded via Blockchain, thus creating blockchains that guarantee data immutability. Because it is a database structured in Blockchain, BigchainDB provided the STEPES-BD project with a decentralized and secure architecture system, ensuring its scalability even beyond a simple PoC. This makes it possible to use a wide network, with several nodes taking full advantage of the consensus algorithms used in BigchainDB, an intrinsic characteristic of a network supported by Blockchain technology. Another relevant factor found was, because it is a database structured in Blockchain, with its core functioning in the MongoDB database, the query tools available in MongoDB’s documentation have become important to facilitate the location of relevant information. It is important to emphasize that data manipulation to insert new information or modifications directly into the MongoDB database is not allowed. This is because all data changes in the Collections used in the Blockchain must be managed by the internal algorithms at the service of BigchainDB and, if there was direct data manipulation in MongoDB, there will also be a change in the identification hash of the block that will contain this record. As a result, a blockchain violation will occur. These facts, traceable by the mechanisms of BigchainDB, are prevented through the consensus algorithms obtained in the implementation of a distributed system.
27.3.2 The Infrastructure for the STEPES-BD Project The infrastructure used, for example by the Patient and Physician STs, for the outlined architecture, was built on a low-cost dedicated server composed of: an Intel I5 2.5 Ghz processor; 8 GiB of RAM; 500 GiB HDD; with the Ubuntu 18 Operating System; and responding on the Internet.
222
M. A. M. da Conceicao et al.
Fig. 27.1 The 3-Tier Architecture of the STEPES-BD project
Fig. 27.2 The Data Interoperability in the STEPES-BD project
In addition, software layers were created in the Docker Container with MySQL and InfluxDB, and the APIs were developed with Python, using the micro framework Flask. The STs of the Hospital and Supplier Segments used an infrastructure based on Platform as a Service (PaaS) with the services of Heroku and Python anywhere, to host services related to the segments. In both STs, the Google Cloud platform was used for storage in an integrated relational database.
27.3.3 The Software Application In the front end layer, several integrated technologies were used to assist in meeting the planned requirements. To simulate a complex application environment, where each consumer of the STEPES-BD project system (Hospitals, Doctors, Patients, Suppliers, or Public Agencies) could have the freedom to develop for their own business needs, we sought the use of data with guaranteed interoperability in the use of the RESTful API. In this sense, there was flexibility
27 An Information Management System for the COVID-19 Pandemic Using Blockchain
in the use of technologies by the development teams to build the front end and back end of their applications. Another factor that influenced the adoption of the technologies used was the diversity of skills of the team members, making interoperability the best solution so that there was not a big learning curve on homogeneous technologies for all teams and the project could be completed on fixed time. For the back end creation to the communication between the buses of the business layer and the interface layer, the following technologies were used: Python, for the development of business rules; together with Django, for communication with the application layer. Technologies such as NodeJS, Express, and Sequelize were also used to create the back end of the Hospital follow-up and provided the integration between the Business Layer and the Presentation Layer, through technologies such as Handlebars and Bootstrap.
27.3.4 The Data Interface For the data persistence layer, in order to guarantee data integrity, reliability, and security, Relational (SQL) and NonRelational (NoSQL) database technologies were used. Among them, the following stand out: MySQL, for record storage of domain entities in the Patient, Physician, Hospital, and Supplier segments; and BigchainDB, for medical care record stores and hospital supply purchases. The BigchainDB, in addition to its technological base, contains distributed Big Data architecture inherited from the MongoDB technology (NoSQL), and it also has, as additional features, the following characteristics provided by Blockchain technology: decentralized control; immutability of data; and transferring digital assets. The InfluxDB was used to store vital information (heart rate, blood pressure, and body temperature) of Patients monitored at a distance during the quarantine.
27.3.5 Testing the Application In parallel with the development of the STEPES-BD project, the course CE-229 Software Testing was also taught. In it, students were introduced to the needs for software testing and the types, techniques, examples, and existing and updated tools for the elaboration of Test Cases for the Project, using the Test Driven Development (TDD). TDD is a software development technique related to the concepts of verification and validation. This technique is based on a short cycle of repetitions, where developers initially write an automated test case and/or define desired improvements as new system functionality. Then, a code is produced to be validated in test cases, for later restructuring
223
and/or refactoring, using acceptable standards and transforming USs into Project functionalities. The State Transition Test was one of the types of tests used in this Project. It was used to analyze and validate the behavior of the developed System Prototype, concerning the defined scope and the fulfillment of the specified requirements. A tool used to carry out this type of test was the State Transition Diagram, used to model certain discrete behaviors of the system, involving an initial state, states traveled, exit points, transitions defining paths to go between states and a final state of termination of the entire testing process. Because this Project was based on Big Data, it became necessary to carry out some stress tests to predict how the system would behave in the face of a large sets. Another type of test also used in this Project was Cloud Testing, remotely mobilizing various computing services, in terms of hardware and software. In this type of test, at least three essential services were tested for performance, security, and functionality, involving Hardware, Infrastructure, and Software: PaaS – Platform as a Service; IaaS – Infrastructure as a Service; and SaaS Software as a Service. In this Project, the Pairwise Testing technique was also applied, coming from mathematical bases where, through orthogonal matrices, it became possible to stipulate fewer numbers of tests to be performed, ensuring better results when performing only the combinations of tests possible.
27.4
The Proof of Concept (PoC)
This section describes the functionalities of the STEPESBD computer system presented as deliverables performed incrementally over its four Sprints The Sprint Zero: This Sprint was dedicated to the identification of existing training and to the leveling of knowledge among students. On this Sprint zero, students took several extra-curricular courses, free of charge, via Internet. In addition to identifying the main skills of the students, initial guidelines were given for the formation of the Teams, defining the Scrum Roles to be performed, in order to allocate all members of the three Courses CE-245, CE240, and CE-229 into the four STs of: Patients, Physicians, Hospitals, and Suppliers. The Sprint 1: In this Sprint, the Scrum Teams developed the functionalities so that it was possible to manage the registration data of Patients, Physicians, Hospitals, and Suppliers. Before, however, it was necessary to carry out a survey on data to be persisted and to develop data models. This Sprint considered the following deliverables: the conceptual models and the entity-relationship models for all four segments; the development of the back end and the front end of the Registration Screens for all four segments; the
224
creation of servers to support the computational system such as: web server; database; and the elaboration of test case for 33 artifacts. The Sprint 2: This Sprint had three main objectives: to start the inter-team integration; to develop the pre-service process, via Web and remote Patient monitoring process; and to use the Blockchain technology to provide traceability and immutability in the transfer process of hospital supplies. In this Sprint 2, the STs Patient and Physician started the integration of their segments, through the development of the Remote Pre-service functionality, in which Patients inform their symptoms and the System indicates if a Patient must attend a Hospital in person, to avoid unnecessary travel to Hospitals during the pandemic period. In case of need to travel to a Hospital, a Patient will be able to choose, at a distance and via Internet, the most convenient date and time for a face-to-face consultation. At the end of the procedure in the Face-to-face attendance, the STEPES-BD system allows the generation of a Medical Report with its digital signature, using its identification keys provided by Blockchain technology. To verify the use of the Blockchain technology in this Project, the concept of ownership over Assets was implemented. In this case, the Asset considered was a Medical Report. When creating this Asset, a Physician performs, via Blockchain, a Transaction of the CREATE type, in which the Medical Record is effectively inserted into the Blockchain. At the end of an appointment, a physician transfers possession of the record to the Patient, characterizing the practical application of the Medical Record concept. With this, it is possible to prove in this PoC the realization of Transactions of the type TRANSFER from Physicians to Patients, thus completing the implementation of a Face-to-face Care Process in the STEPES-BD project system via Blockchain. After Patients have attended the consultations in person, they were able to be sent to home quarantines in situations where the symptoms were of low or medium severity.
Fig. 27.3 The transfer of hospital supplies using Blockchain technology
M. A. M. da Conceicao et al.
Also in this Sprint 2, the members of the STs were able to implement a Remote Monitoring Process in real time, containing vital information for this ST of Patients. In this case, remote monitoring of disease evolution in Patients became possible, collecting data on heartbeat, blood pressure, and body temperature, via Internet applications. The Remote Monitoring functionality was implemented using the InfluxDB Time Series Database and viewed through the Grafana application. However, for data generation, no real Internet of Things (IoT) devices was used, because implementations in hardware devices were out of the scope of this Project. In this Sprint 2, the members of the Hospital ST and Supplier ST were also able to perform integration through the Resource View. In this case, Hospital can request from Supplier new resources such as equipments, medical supplies, and medicines, and this transfer of resources can also be registered, via Blockchain, providing guarantees of immutability and traceability. In this case, access to the Blockchain was also accomplished through the BigChainDB tool, which abstracted all the complexity of managing the blockchain, as shown in Fig. 27.3. In addition, the ST of the Hospital segment also implemented a register of hospital beds, so that it was possible to know, in real-time, the situation of hospital beds in the city of São José dos Campos, to support decisions of the municipal public authorities. The Sprint 3: The main objectives of this Sprint 3 were: to carry out the integration between all the STs; to generate a large data set of fictitious data to justify the use of Big Data technologies; to provide records of face-to-face assistance; to implement and demonstrate Dashboards involving contagion maps that can be updated in realtime; to provide the creation of visualizations, assisting managers in their decision making; and to promote the exploration and implementation of EHR concepts with the traceability and immutability of Blockchain technology.
27 An Information Management System for the COVID-19 Pandemic Using Blockchain
To achieve the objectives proposed for this Sprint 3, the main deliveries made by the members of the four STs were simulated and loaded with 1M of Fictitious Patient Records and 2,000 Physician Records. With that, it was possible to generate data to update a Dashboard, synthesizing information through procedures implemented using the Spark tool. For Proof of Concept (PoC) purposes, during the implementation of the Face-to-Face Attendance functionality, a Physician selected a Patient from his/her agenda for a Consultation, then he wrote a diagnosis in a textual field. Based on pre-programmed options, the system made it possible for the physician to opt for the Hospitalization of the Patient in a given Hospital, and to request and register all the disposable items needed to be provided by certain Suppliers and used in a Consultation, including for example, tests of the COVID-19 virus, masks, disposable gloves, among others. In this case, the data related to Face-to-face care involving Patients, Physicians, Hospitals, and Suppliers of materials and medicines were also consolidated through the use of Blockchain. The Dashboard built in the Front End provided a visualization of the panorama of the COVID-19 Pandemic and made it possible to identify the critical regions of the city of São José dos Campos, with the highest incidences of those infected, as shown in Fig. 27.4. In addition, some indicators shown in Fig. 27.5 were also implemented, to provide an overview of the COVID-19 Pandemic in the city. The front end containing the indicators
Fig. 27.4 Simulated contagion map of the city of São José dos Campos
225
was written in React and it consumed data provided by Python APIs that was consolidated from the Patient and Physician segments, and also in NodeJS consolidated from Blockchain. Also in this Sprint 3, the integration of functionalities between Online Pre-Attendance and Face-to-face Attendance provided the effectiveness of the initial proposal for Medical Care Records Stored in the Blockchain. This integration made it possible to track any relevant information in medical care because, through public and private keys and hash mechanisms, it became possible to identify those involved, related medicines, participating suppliers, and the attendance location. Thus, it was ensured that all documents generated by the Prototype of the System could be certified with digital signatures, such as the Fictitious Medical Record and the Fictitious Death Certificate, as shown in Fig. 27.6. In addition, the Electronic Death Certificate shown in Fig. 27.6 could also be implemented and issued right after the confirmation of the death of a Patient, avoiding family stresses, possibilities of fraud, or delays in communications to public authorities.
27.5
Conclusion
This article described de development of a Management Information System for the COVID-19 Pandemic using Big
226
M. A. M. da Conceicao et al.
Fig. 27.5 Dashboard indicators consolidated from the physician and hospital segments and Blockchain
Fig. 27.6 An example of death certificate digitally signed.
Data, Blockchain, and other emerging technologies applied to the STEPES-BD project. The use of these emerging concepts and technologies has provided students with the opportunity to deepen their practice in topics theoretically learned from the courses of: CE245 Information Technologies; CE-240 Database Systems Project; and CE-229 Software Testing. At the end of the 1st Semester of 2020, students presented the results of the STEPES-BD Project as a Proof of Concept to professors and guests from the industrial and academic sector in the city of São José dos Campos. The developers of this Project believe that the application of several emerging technologies aimed at information
management to help combat the COVID-19 Pandemic, as a real problem that afflicts the whole world, represents an important opportunity for updates and transfers of experiences and information between academia and society. At the end of this report involving the development of a prototype of an academic system as a Proof of Concept, the authors of this article recommend that the technologies used in the next Projects be defined, preferably, in Sprint 0, to provide better theoretical use and practical application by members of the Scrum Teams (STs). It is also recommended that public health stakeholders should be involved in the next Projects, aiming at providing
27 An Information Management System for the COVID-19 Pandemic Using Blockchain
technological transfers and generating greater benefits to society with academic projects like this one. In the next Projects, the use of the Cloudera Cluster is also recommended, which is at an advanced stage of implementation in the Computer Science Division of ITA, to support even more scalable experiments with larger data sets. It is believed that this academic alternative is viable, instead of using commercial clusters, such as Amazon Web Servers (AWS), among others, which will make it possible to expand Proofs of Concepts and new and emerging Research and Development (P\&D) involving Information Technologies (ITs) such as Big Data, Blockchain, among others. As the base of the project is a Document-oriented database Health Level Seven (HL7) or other international standards weren’t a worry in this Proof of Concept because NoSQL databases offer flexible queries and other resources and also give a good answer for scaling big types of files like CT Scans, MRI and others converting them to Base64 type. However, BigchainDB makes the perfect union of big data distributed database and blockchain characteristics (decentralized control and immutability) not preventing cyber attacks but it is enough to guarantee a high level of security by using Interledger Protocol (ILP) and its specific conditions configurations like the minimum validators necessary for a transaction could be processed that may be set up. For future work, the authors suggest the natural continuation of the STEPES-BD project, integrating Patient monitoring through sensors capable of measuring, in real-time, heartbeat data, blood oxygenation, among others. It is also suggested, for future work, the creation and development of appropriate policies for accessing the STEPES-
227
BD Project or similar ones to provide visualizations of general data and also specific information about Patients, Physicians, Hospitals, and Suppliers to better assist managers of the Health Municipal, State, Federal, and other stakeholders. Acknowledgments The authors of this article would like to thank the Brazilian Air Force, the Brazilian Aeronautics Institute of Technology, the Ecossistema Negocios Digitais Ltda (WiBOO/Wibx), and the Casimiro Montenegro Filho Foundation, for the support and infrastructure offered during the development of this ALFA Project (MEC-ITA).
References 1. J. Kent, Could covid-19 help refine ai, data analytics in healthcare. Accessed 24 June 2020 [Online]. Available https:// healthitanalytics.com/features/could-covid-19-help-refine-ai-dataanalytics-in-healthcare 2. P. Mirchandani, Health care supply chains: Covid-19 challenges and pressing actions. Accessed 06 May 2020 [Online]. Available https:/ /www.acpjournals.org/doi/10.7326/M20-1326 3. P.K. Sinha, G. Sunder, P. Bendale, M. Mantri, A. Dande, Electronic Health Record: Standards, Coding Systems, Frameworks, and Infrastructures (Wiley, 2012) 4. J. Sutherland, A. Viktorov, J. Blount, N. Puntikov, Distributed scrum: Agile project management with outsourced development teams, in 2007 40th Annual Hawaii International Conference on System Sciences (HICSS’07). IEEE, 2007, pp. 274a–274a 5. B.W. Tuckman, Developmental sequence in small groups. Psychol. Bull. 63(6), 384 (1965) 6. T.C. Shan, W.W. Hua, Solution architecture for n-tier applications, in 2006 IEEE International Conference on Services Computing (SCC’06). (IEEE, 2006), pp. 349–356 7. Key concepts of BigchainDB. Accessed 13 July 2020 [Online]. Available https://www.bigchaindb.com/developers/guide/keyconcepts-of-bigchaindb/
Machine Learning for Classification of Cancer Dataset for Gene Mutation Based Treatment
28
Jai Santosh Mandava, Abhishek Verma, Fulya Kocaman, Marian Sorin Nistor, Doina Bein, and Stefan Pickl
Abstract
The objective of this paper is to develop a Machine learning model that can classify cancer patients. Gene mutation-based treatment has a very good success ratio, but only a few cancer institutes follow it. This research uses natural language processing techniques to remove unwanted text and convert the categorical data into numerical data using response coded and one-hot encoded. Then we apply various classification algorithms to classify the training data. The proposed system has the advantage of reducing the time to analyze and classify clinical data of patients, which translates into less wait time for patients in order to get results from pathologists. The results of our experiment will demonstrate that the Stacking Classifier algorithm with One-Hot encoding and Term Frequency – Inverse Document Frequency (TF-IDF) techniques perform better than other Machine Learning methods with around 67% accuracy on the test data.
Keywords
Gene mutation · Cancer patient · Machine learning · ML · Text data classification · Natural language J. S. Mandava · F. Kocaman · D. Bein () Department of Computer Science, California State University, Fullerton, Fullerton, CA, USA e-mail: [email protected]; [email protected]; [email protected] A. Verma Department of Computer Science, California State University, Northridge, Northridge, CA, USA e-mail: [email protected] M. S. Nistor · S. Pickl Department of Computer Science, Universität der Bundeswehr München, Neubiberg, Germany e-mail: [email protected]; [email protected]
processing · NLP · Random forest classifier · Stacking classifier · Personalized medicine
28.1
Introduction
A sizable set of people suffer from various types of cancer every year, and nearly half of them die without proper diagnosis. According to the National Cancer Institute [1], gene mutations will enhance the treatment of disease by supplying effective medicines and decreasing the number of deaths. A considerable amount of scientific evidence collected over the last decades can be used to enhance cancer detection. Data on the health of past patients with diagnostic medicine are found in DNA or gene mutation. Based on the previous results, this approach can help physicians to treat future patients. Gene mutations are carried out by manual techniques. An experienced health care professional has to manually apply knowledge to track specifics and suit the gene of the patient. Using machine learning algorithms, the clinical data analysis can be automated to improve the diagnosis and treatment accuracy and time. Artificial Intelligence can predict the gene mutation in the patient groups by implementing machinelearning algorithms. In gene mutations linked to lung cancer, family genetics could play a significant role in analyzing the likelihood of cancer risk [2]. While gene mutations are known for lung cancer diagnosis, they can be used to collect the data and compare them manually. However, the alternative is to replace it with greater precision through the use of machine learning algorithms. This research applies various natural language processing algorithms to the clinical dataset collected from lung cancer patients to perform data cleaning. Furthermore, the paper focuses on applying machine learning classification. According to Kourou et al. [3], “Machine Learning (ML) Algorithms can detect and classify trends and
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_28
229
230
J. S. Mandava et al.
interrelationships from complex data sets when effectively predicting future results from cancer.” The preparation, testing, and prediction are three phases of ML. Algorithms do software cleaning, replication, redundancy, and data preparation, whereas the ML techniques are used in the validation process for test data using specialized expertise. To ensure the correct algorithm is chosen, the accuracy should be determined. Next, the algorithm is used for the prediction of new data. The proposed system was implemented in the Jupyter notebook using Python language. The problem encountered in the gene mutation treatment is that any manual process takes more time to analyze the patient’s data than an automated one. For the past several years, there has been much improvement in machine learning algorithms, parallel processing, and the ability of algorithms to process data at scale. We implemented machine learning algorithms on the dataset provided by Memorial Sloan Kettering Cancer Center [4], which has around 3200 data with the clinical text of each patient. We analyze two separate files, described as follows: • Training variants – the file that includes the definition of genetic variations used for training. Training variants Fields are IDs (including the gene ID for the row of clinical evidence used to link the mutation), gene (gene where the genetic mutation is located), variations, class (includes 1– 9 classes). • Training text – contains the clinical text of the patients with their ID. The proposed system has the following advantages. First, it reduces the time to analyze and classify the clinical data of patients, so this translates into less waiting time for patients to get results from pathologists. Second, our model can handle a large amount of data at a time. The paper is organized as follows. In Sect. 28.2, we present background information, followed by the proposed system model in Sect. 28.3. Simulation results are shown in Sect. 28.4, concluding remarks and future work are presented in Sect. 28.5.
28.2
Background
In the eighteenth century, cancer was treated using surgery, which was considered then as primary treatment. Postsurgery herbal therapy, castor oil, and arsenic were administered to the patient. Radiation therapy came into existence in 1895 but only cured a few cancer types [5]. In the treatment of lymphoma cancer with the subsequent
approval of the FDA, immunotherapy was introduced at the end of 1987. In 1990, gene-based treatment was approved by the FDA for immunodeficiency disorder. Later, gene-based treatment was used on cancer patients and successfully mitigated different types of cancer such as brain tumors, acute lymphocytic, and others. According to Memorial Sloan Kettering Cancer Center [4], genetic tests are conducted in various cancer types and are divided into nine different classes for treatment. Each class is a cluster of gene-related treatments in which patients get treated according to the class. Patient classification happens through three phases, and eventually, the appropriate patient class is determined. The first phase is where a collection of the patient’s records is used, such as PET scan, CT scan, and previous health reports. Then pathologist looks for the clinical evidence from the medical literature and converts it into clinical text for further analysis. In the last phase, pathologists classify the patient into nine different classes by analyzing the clinical text. The final phase usually takes one to three days to classify the patients into a particular category. Lung cancer is a type of cancer that can be traced from family history. Although according to Memorial Sloan Kettering Cancer Center [4], many people do not get proper treatment because the selected medication does not work on them, this is where gene mutation-based treatment can be helpful. The use of personalized precision medicine improves patients’ survival and quality of life [6]. However, the patient’s clinical data needs to be analyzed manually, and that requires the time of experienced staff to make the decisions. The authors in [7] have used the Cancer Genomic Atlas results, including patients who have lung adenocarcinoma (LUAD). In order to acquire genes for LUAD patients, the author applies a machine learning algorithm. Further information on the dataset is given by using the KNN algorithm, decision trees, SVM, and Naïve Bayes. They show that ZNF560 and DRD3 are the positive genes capable of surviving LUAD. In order to assess the patient survival rate, the top six genes are later tested by the research model. The findings have been comparable to the original. The authors in [8] implemented deep learning algorithms to classify lung cancer into gene mutations. They considered two major lung cancer types: Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LUSC). According to the report, conventional neural networks are used to study the images and classify the genes. After training the model, they researched in the real world through means of new data of the patient for testing. The results were compared to the results of the pathologist from different sources, which show there is slight misclassification of the model. Their research does not involve other types of lung cancer and focuses on only two major types.
28 Machine Learning for Classification of Cancer Dataset for Gene Mutation Based Treatment
231
Fig. 28.1 System architecture Table 28.1 Response encoded table Gene ALK CBL KIT
28.3
Class 0 4/20 14/35 4/10
Class 1 5/20 10/35 3/10
Class 2 11/20 11/35 3/10
Proposed System Architecture
Our proposed system architecture is shown in Fig. 28.1. We implemented our program in Jupyter Notebook using Python language. Unfortunately, the final phase of the gene mutation-based process takes more time, so we replace that with the machine learning algorithm. We have approximately 3200 data points [4] divided into three parts: training data, cross-validation data, and test data. Before dividing the data, we need to clean the data and remove empty rows from the data. Using Natural Language Processing (NLP), we removed unwanted text in the clinical data of the patients. After splitting data, we implemented Response Coding and One-Hot encoding to convert the text data to numerical form for fitting the model in machine learning algorithms. Then the genetic mutations were classified into one of nine classes. There are two columns of categorical data, which are genes and variations. To categorize the data, we implement one-hot encoding and response coding. One-hot encoding is
implemented using the count vectorizer in the scikit-learn library. Response coding is part of the machine learning technique used for categorizing the data. This technique is implemented to represent the data point by calculating the probability for each class by category. For example, suppose we have 233 unique genes for nine classes. We get 233 features by calculating the probability for each class. Suppose we take the three genes as three classes in the data. Table 28.1 describes how response coding is done for this example. There are three types of genes such as ALK, CBL, and KIT. There are twenty types of patients with genes as ALK, and twenty have different classes. For example, in Table 28.1, Class 0 has 4 ALK genes. The probability is calculated by the total number of genes in the class, divided by the total number of the specific gene contained in the data. According to [9], combining Natural Language Processing with Machine Learning algorithms helps search for patterns in text data. We implemented both Response Coding and scikit-learn’s TF-IDF vectorizer methods for the text data. We then combine the features using One-Hot encoding on gene and variations with TF-IDF on the text data in each data set separately. We also combined all three features, where Response Coding was applied to gene, variation, and text data. The following classifiers with different NLP techniques were tested with different values using the CalibratedClassifierCV from the scikit-learn library. The best hyperparameters based on the smallest log-loss for each classifier were found during cross-validation. We then implemented
232
J. S. Mandava et al.
the following classifiers using the best hyperparameters for each one. We use the following classifiers: • Naïve Bayes Classifier is based on Bayes theorem for classification technique and assumes that predictors are independent. Thus, the existence of a specific feature in a class is irrelevant to any other feature in Naive Bayes classification. MultinomialNB is implemented in this research, and it can be imported from the scikit-learn library. • K – Nearest Neighbors (KNN) can be used for statistical problems of classification as well as regression. • Linear Support Vector Machine (SVM) is a linear model of a supervised machine learning algorithm which can be implemented in regression and classification modes. This is used primarily in the classification model. The value of each data element is the value of each coordinate in the SVM algorithm as a point in the n-dimensional field. Then we define the hyperplane, which differentiates between the two groups very well. • Random Forest Classifier is a guided algorithm of learning. It can be used both for regression and classification. It is based on the idea of bagging. Many weak classifiers named decision trees pooled together are better than a single strong classifier. The random forest generates randomly selected data samples for decision trees, obtains predictions of each tree, and selects by voting the best solution. It also gives a very clear indication of the value of the feature. • Stacking Classifier is a technique for combining many classification models with a meta-classifier. The individual classification models are trained based on the entire training set. The meta-classifier is then installed on the output from the individual model classifications within the ensemble-meta-features. The meta-classifier may also be educated on the predicted class labels or ensemble probabilities.
28.4
Experimental Results
We implemented various machine learning algorithms on the dataset provided in [4], which has around 3200 data with the clinical text of each patient. Our results are displayed in Fig. 28.2. Best accuracy is obtained from the random forest and stacking classifier. Table 28.2 shows each model’s misclassification accuracy and log-loss on the test data we use in this project.
Table 28.2 Classification results ML classifier Naïve Bayes with One-Hot Encoding and TF-IDF Naïve Bayes with Response Coding KNN with One-Hot Encoding and TF-IDF KNN with Response Coding Logistic Regression (Balanced) Logistic Regression (unbalanced) Linear SVM(Balanced) Random Forest Stacking Classifier
28.5
Mis-classification test accuracy 38.25%
Test log-loss 1.2939
60.09%
1.7293
39.91%
1.2935
47.74%
1.5251
34.64%
1.0891
34.49%
1.0850
35.84%
1.2016
33.58% 33.37%
1.0831 1.0183
Conclusion and Future Work
In this paper, we describe Machine learning models that can classify cancer data of the patients. Our approach focuses on replacing the final phase of the gene mutation treatment for cancer using a machine learning algorithm. This implementation helps in improving the accuracy of the analysis and reduces the wait time for classification. The classification was performed using several popular machine learning algorithms. The test log-loss of the Stacking classifier was close to 1. The classification accuracy came out to be about 67% on test data. The problem is reasonably challenging, and more work is needed to reduce the error rate of the model so that the model accuracy can match the real-world accuracy of experienced pathologists. Future work for the project will mainly focus on improving the accuracy of the model. In addition, we would expand the dataset beyond 3200 data points to implement deep learning algorithms that require larger amounts of training data and may produce improved classification accuracy. Furthermore, such a model can also be adapted to a triage process. For example, it might reduce the time to analyze and classify the triage protocols and the exiting clinical data of patients to improve the treatment response. Acknowledgment This research was sponsored by the NATO Science for Peace and Security Programme under grant SPS MYP G5700.\def\secrule{\color{hfblkcolor}\rule{\hsize}{1pt}}
28 Machine Learning for Classification of Cancer Dataset for Gene Mutation Based Treatment
233
Fig. 28.2 Comparison of log-loss performance for various machine learning classification algorithms
References 1. National Institute of Cancer, The Genetics of Cancer. October 12, 2017. Available online at https://www.cancer.gov/about-cancer/ causes-prevention/genetics. Last accessed 10 Dec 2020 2. Genetics Home Reference, (2019). Available online at https:/ /ghr.nlm.nih.gov/condition/lung-cancer#sourcesforpage. Last accessed 10 Dec 2020 3. K. Kourou, T.P. Exarchos, K.P. Exarchos, M.V. Karamouzis, D.I. Fotiadis, Machine learning applications in cancer prognosis and prediction. Comput. Struct. Biotechnol. J 13, 8–17., ISSN 20010370 (2015) 4. Memorial Sloan Kettering Cancer Center, MSK-IMPACT: A Targeted Test for Mutations in Both Rare and Common Cancers. Available online at https://www.mskcc.org/msk-impact. Last accessed 10 Dec 2020 5. M.H. Amer, Gene therapy for cancer: Present status and future perspective. Mol. Cell. Ther. 2, 27 (2014). https://doi.org/10.1186/ 2052-8426-2-27
6. F.B. Marques, G.F. Leal, G.N. Bettoni, O.N. de Souza, Integration of bioinformatics and clinical data to personalized precision medicine, ITNG 2021 18th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing, vol 1346 (Springer, Cham). https://doi.org/10.1007/ 978-3-030-70416-2_23 7. H.-J. Cho, S. Lee, Y.G. Ji, D.H. Lee, Association of specific gene mutations derived from machine learning with survival in lung adenocarcinoma. PLOS ONE 13(11), e0207204 (2018). https://doi.org/ 10.1371/journal.pone.0207204 8. N. Coudray et al., Classification and mutation prediction from non– small cell lung cancer histopathology images using deep learning. Nat. Med. 24(10), 1559–1567 (2018) 9. C. Vielma, A. Verma, D. Bein, Single and multibranch CNNbidirectional LSTM for IMDb sentiment analysis, in ITNG 2021 17th International Conference on Information Technology-New Generations, Advances in intelligent systems and computing, vol. 1134, (Springer, Cham, 2020). https://doi.org/10.1007/978-3-03043020-7_53
Part VI Machine Learning
Performance Comparison Between Deep Learning and Machine Learning Models for Gene Mutation-Based Text Classification of Cancer
29
Fulya Kocaman, Stefan Pickl, Doina Bein, and Marian Sorin Nistor
Abstract
Identifying genetic mutations that contribute to cancer tumors is the key to diagnosing cancer and finding specific gene mutation-based treatment. It is a very challenging problem and a time-consuming job. Currently, clinical pathologists classify cancer manually, and they need to analyze and organize every single genetic mutation in cancer tumors from clinical text. The text data analysis can be automated using Deep Learning and Machine Learning classification techniques to ease the manual work needed to extract information from clinical text. This paper aims to analyze the performance of Machine Learning and Deep Learning methods to classify cancer from gene mutationbased clinical text data. This paper uses Natural Language Processing techniques, namely, CountVectorizer and TfidfVectorizer, and Keras API’s One-Hot encoding and tocategorical utility, to vectorize the categorical and text data and transform them into numerical vectors. Machine Learning classification algorithms and Deep Learning methods are then applied to the extracted features, and the most accurate combination of feature extraction and a classifier is discovered. Keras API’s Embedding Layer (Word Embeddings) and Bidirectional Long-Term ShortTerm Memory (Bidirectional LSTM) techniques using original and augmented text data from NLPAug library are applied as Deep Learning methods. The Keras Word Embeddings using augmented text data ha performed the highest with an accuracy of 80.67%, the weighted average F. Kocaman · D. Bein () Department of Computer Science, California State University, Fullerton, Fullerton, CA, USA e-mail: [email protected]; [email protected] S. Pickl · M. S. Nistor Department of Computer Science, Universität der Bundeswehr München, Neubiberg, Germany e-mail: [email protected]; [email protected]
precision of 0.81, recall of 0.81, F1 score of 0.81, and the log loss of 0.6391. As for the Machine Learning classification algorithms, Random Forest and Stacking classifiers are explored within this paper, and the highest accuracy of 67.02% is achieved from the Random Forest classifier with the weighted average precision of 0.70, recall of 0.67, F1 score of 0.65, and the log loss of 1.0523. Keywords
Machine Learning · Random Forest · Deep Learning · Embedding Layer · Keras API · Text data classification · Gene mutation · NLP · NLPAug · Personalized medicine
29.1
Introduction
Cancer is one of the primary causes of death globally, and every year millions of people die of cancer-related diseases [1]. Specific alterations in protein-carrying genes can lead to abnormal cell growth and become cancer. These genetic mutations could damage the DNA of a cancer patient, and consequently, their descendants may inherit a similar type of cancer. Therefore, these gene mutations are used in targeted treatments for practical medicine. For example, targeted treatments would currently work on specific gene mutations linked to non-small cell lung cancer with EGFR, KRAS, ALK, and other gene mutations [2]. With the help of DNA sequencing and bioinformatics analysis methods, doctors and researchers can identify specific genetic mutations linked to certain types of cancer for early diagnosis and personalized treatments that might work best for their patients. This experiment uses Count Vectorizer, and TF-IDF from the scikit-learn library and Keras API’s One-Hot encoding and to-categorical utility to vectorize the categorical and text
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_29
237
238
data and transform them into numerical vectors. We then analyze and identify genetic mutations from text-based clinical literature using Deep Learning and Machine Learning classification methods. We built Deep Neural Networks using Embedding Layer and Bidirectional LSTM using original and augmented text data in Keras API. We used text augmentation using the NLPAug library with Bidirectional Encoder Representations from Transformers (BERT) Contextual Embeddings. As for the Machine Learning classifiers, we used Random Forest and Stacking Classifiers. The dataset provided from Memorial Sloan Kettering Cancer Center (MSKCC), one of the most extensive private cancer treatment and research institutes in New York, is used in this paper. MSKCC has provided an expert-annotated precision oncology knowledge base with thousands of mutations manually annotated by world-class researchers and oncologists for studying gene classification. The dataset contains 3321 patients’ data and has the clinical text of each patient. Following describes the two files used in this paper: • The patients’ data (training/test variants) with their IDs contain Gene, Variation, and Class (1–9 classes). • The clinical text (training/test text) of the patients with their IDs provides the clinical evidence that researchers and oncologists use to classify the genetic mutation. The classification methods are evaluated using classification accuracy and log-loss. Even though this problem can be considered a text classification, it is more complex and challenging than other traditional text classifications because each clinical text data entry contains very long and detailed medical terminologies. Moreover, some samples share duplicate data entry, while having different class labels [3]. Figure 29.1 also depicts how unbalanced the distribution of classes is over the entire data set.
Fig. 29.1 Distribution of classes
F. Kocaman et al.
The paper is organized as follows: Background information is presented in Sect. 29.2. Then, Sect. 29.3 describes the proposed system architecture, while Sect. 29.4 shows the results of our experiment. Finally, Sect. 29.5 concludes this paper with a conclusion and future work.
29.2
Background
For the last several years, there have been considerable improvements in genetic testing, which detects cancer from mutations in specific genes. The genetic test called DNA sequencing test can read DNA and compare the sequence of DNA in cancer cells with that in normal cells such as blood samples to identify genetic changes in cancer cells. It can also reveal inherited genetic mutations associated with an increased risk of cancer [4]. According to Memorial Sloan Kettering Cancer Center [5], many people do not get proper treatment because the selected medication does not work on them; this is where gene mutation-based treatment can be helpful. The data from DNA sequencing helps doctors make a proper prognosis, and it offers them targeted treatment options to increase the survival rate of their patients. However, there are thousands of genetic mutations of a sequenced cancer tumor, and these mutations need to be differentiated from neutral mutations [6]. The classification process is challenging and time-consuming because a clinical pathologist manually analyzes these genetic mutations from clinical text. Like many other fields, the medicine and health industry relies on developing the latest technology to improve the accuracy of digital techniques. Correctly diagnosing a patient and prescribing the proper treatment are the essential parts of the medicine domain to cure diseases consequently [7]. Therefore, there is a growing trend of using Deep Learning and Machine Learning to assist medical personnel in predicting and detecting cancer [8]. Machine Learning classification techniques can ease the manual work needed to extract information from medical records and clinical text on unwanted, unorganized, and multi-dimensional data [7]. Applying Deep Learning and Machine Learning techniques have recently become a widespread application in cancer diagnosis and precision medicine for gene mutation-based treatments. It is mainly due to the improvements in hardware capabilities and cloud computing to store vast quantities of data and software capabilities to efficiently manage and process big data in the last several years. Therefore, an automated clinical data analysis can improve the diagnosis and treatment in terms of accuracy and time using Machine Learning algorithms. Furthermore, an automatic Machine Learning Algorithm implementation would also decrease the mistakes caused by classifying mutations manually and further support the medical domain.
29 Performance Comparison Between Deep Learning and Machine Learning Models for Gene Mutation-Based Text…
Li and Yao [9] used Term Frequency-Inverse Document Frequency (TF-IDF) as a Natural Language Processing (NLP) technique for feature extraction and conversion and label encoding to convert genes and their variations into feature vectors for training to employ normalized non-linear features. They implemented Extreme Gradient Boosting (XGBoost) and Support Vector Machine (SVM) for classification on the dataset from Memorial Sloan Kettering Cancer Center (MSKCC) through the Kaggle [6] competition. The result of their experiment showed that XGBoost outperformed SVM and produced better predictive results. Furthermore, their investigation revealed that XGBoost outperformed SVM with an accuracy of 66%. Along with using TF-IDF to transform the clinical text into a numerical feature vector, Waykole and Thakare [10] used a one-hot encoding technique to extract features from genes and their variations from the same MSKCC dataset. They applied a basic Logistic Regression technique to the extracted features and achieved an accuracy of 64%. They also emphasized how difficult it is to classify genetic mutations based on medical literature for human experts due to decoding clinical evidence. The authors in [11] experimented with many different Machine Learning classifiers in their study of predicting and early detecting diseases. After the feature selection, they used Logistic Regression with the selected feature, SVM with a linear kernel for their analysis to classify data to maximize the margin between the hyperplane, AdaBoost, Random Forest, and Decision Tree classifiers. They concluded that Logistic Regression performed on the heart dataset had the highest accuracy, SVM performed superior to others in their diabetes prediction, and AdaBoost classifier performed the best for breast cancer detection among the other machine learning algorithms they used. However, they also concluded that the Random Forest and Decision Tree methods did not perform better than other classifiers on three separate datasets. Zhang et al. [3] explored Word2Vec (Skip-Gram), a predictive neural word embedding model that captures the semantic relationship of words in a text from the same MSKCC dataset. They developed an ensemble of nine primary gradient boosting models applied to the combined features, which showed the best performance in their evaluation with an accuracy of 67%. The authors in [12] researched four different Machine Learning classifiers on the MSKCC cancer dataset to automate the process. The Logistic Regression classifier with class balancing and one-hot encoding scheme was found to be the most efficient model with an accuracy of 65.23% in their work. Saxena & Prasad [8] experimented with five Machine Learning algorithms: Logistic Regression, Support Vector Machine, Naïve Bayes, K-Nearest Neighbors, and Maximum Sensitivity Artificial Neural network (ANN) in their paper.
239
They concluded that Maximum Sensitivity ANN using Multilayer Perceptron produced the most accurate results based on less sensitivity variation in predicting and detecting cancer disease.
29.3
Proposed System Architectures
29.3.1 Deep Learning Neural Network Models The following Deep Neural Network models are implemented in Google Colab Pro with P100 GPUs using Python language and Keras API. These models are trained only using the clinical text data to capture the semantic relations in the text. After the initial cleaning, the text transforms before getting fed into a Neural Network (NN). Keras Tokenizer was used to convert the entire text into numeric tensors by tokenizing into tokens, stripping from special string characters, then turning them into lists of integer indices, and finally turning the lists of integers into a 2D integer tensor of shape. We then split the text into training and testing sets and encoded each text data sentence into integer sequences. Moreover, we finally padded the sequences to have the same length and turned the integers sequences into tensors to fit into a NN [13]. We also chose the Panda’s get_dummies function and Keras’ to_categorical utility to vectorize the labels for our following Deep Learning models. After repeatedly modifying the models, training them, and evaluating them on the validation data with different combinations of layers, hyperparameters, activation and loss functions, and batch and epoch sizes, the following Deep Learning networks gave the most optimal configuration for our problem. As the author in [13] suggested, once we developed a satisfactory model configuration, we trained the final model using both training and validation data and evaluated it on the test data. • Embedding Layer uses dense word embeddings to learn from the data. With the help of backpropagation, the Embedding Layer learns a new word embedding at each layer. The Embedding Layer takes the input integer sequences (2D integer tensor) and turns them into embedded sequences (3D float tensor) [13]. Our Embedding Layer network learns 100-dimensional embeddings for each word, turns 2D integer tensor inputs into 3D embedded tensors in the Embedding Layer, flatten the tensor to 2D, and trains two Dense layers with 100 and 9 layers on top for classification. These Dense layers use activation functions of relu and softmax, respectively. In addition, one Dropout function is used to overcome overfitting. The summary of this network is shown in Fig. 29.2.
240
F. Kocaman et al.
Fig. 29.2 Summary of the network used with Embedding Layer
Fig. 29.3 Summary of the network used with Bi-LSTM
• Bidirectional Long-Term Short-Term Memory (Bidirectional LSTM) encodes the input sequence in both states; information from the past and the future states using two independent LSTMs with different parameters. It is known to be helpful in Natural Language processing and aims to improve the performance of one-directional LSTM [13]. Our Bidirectional LSTM network shown in Fig. 29.3 learns 128-dimensional embeddings for each word in the Embedding Layer and trains a Bidirectional LSTM with 64 layers and a Dense layer with 9 layers on top for classification. The Dense layer uses the activation function of softmax, and a Dropout function is implemented to mitigate overfitting. • Text Augmentation using NLPAug library with Bidirectional Encoder Representations from Transformers (BERT) Contextual Embeddings was also explored to handle the overfitting issue with the Deep Learning models. After splitting the entire text data into training and testing sets, we set aside the testing data and augmented some text data using only the training set. During the BERT Augmentation, it substituted 100 words in each text. That data was added to the entire training data frame, and accumulated training text data was shuffled at the end of each text generation [14, 15]. As a result, we augmented around 9000 new text data using only the training data set, which gave us about 11,000 training text data. Figure 29.4 shows the distribution of the text data over each class
after the BERT augmentation. We implemented the same Neural Network structure as the Embedding Layer model presented in Fig. 29.2. After trying out many different configurations of hyperparameters and evaluating with K-fold cross-validation, that model was adequate for our augmented text data set.
29.3.2 Machine Learning Algorithms Figure 29.5 shows the architecture of the following Machine Learning algorithms used in this paper. First, we analyzed the input data using Exploratory Data Analysis to visualize the input data (both data of training variants and clinical text). During preprocessing, the text data is cleaned, unwanted text in the patient’s clinical data is removed. Then, the final data is split into three sets: training, cross-validation, and test data. Features are then extracted using various NLP and word embedding techniques, giving the model a simpler and more focused view of the clinical text. Finally, the extracted features are evaluated against multiple Machine Learning classification algorithms, and the most accurate combination of feature extraction and a classifier is discovered based on their classification accuracies. As a result, the genetic mutations are classified into one of nine classes. The NLP techniques for feature extraction applied in the following Machine Learning algorithms are One-Hot encod-
29 Performance Comparison Between Deep Learning and Machine Learning Models for Gene Mutation-Based Text…
241
classifications within the ensemble-meta-features. The meta-classifier can either be trained on the predicted class labels or probabilities from the ensemble. Our Stacking Classifier is an ensemble of Naive Bayes, Logistic Regression, Linear Support Vector Machine (SVM), and Random Forest classifiers. We first calibrated each classifier using CalibratedClassifierCV with the best hyperparameter values that we discovered previously. As the optimization method for our model, we then implemented the Stochastic Gradient Descent - Classifier (SGD-Classifier) with Logistic regression from the sklearn library using all four calibrated classifiers. Multiple alpha values are fitted in the SGD-Classifier for our model. Moreover, the best alpha for this classifier was chosen by finding the minimum logloss of the cross-validation data set. It was then used to implement our Stacking Classifier model.
Fig. 29.4 Augmented text data distribution over each class
ing using the Count Vectorizer for gene and variation data sets and TF-IDF Vectorizer for text data. These two vectorizers are both from the sklearn library. Once these three features are vectorized on each training, test, and cross-validation data set separately, we combined the three features coming from each data set. • Random Forest Classifier uses the bagging method and creates randomness from a standard dataset. It takes bootstrap samples from the dataset for each tree to create a forest so that the trees are trained on a somewhat different dataset. Each node of the tree gets a random subset of the features, and to add randomness to the decision tree, it can only pick from that subset rather than from the whole set [16]. In our Random Forest classifier model, we implemented the CalibratedClassifierCV using the sigmoid function from the sklearn library to calibrate probabilities in the Random Forest classifier. Multiple copies of different combinations of the number of trees and the depth of each tree are fitted in the CalibratedClassifierCV for our model. The best hyperparameters for this classifier are chosen by finding the minimum log-loss of the cross-validation data set. They are finally used to implement our Random Forest model. • Stacking Classifier is an ensemble learning technique, where combining multiple classifiers is used as a metaclassifier. The models from each classifier are trained based on the entire training set, and the meta-classifier is then installed on the output from the individual model
29.4
Experimental Results
Attempts to automate classifying cancer from genetic mutations using Machine Learning algorithms with the dataset from MSKCC have been tried several times for the last five years. However, Deep Learning methods have rarely been applied to this particular dataset. Therefore, the goal of this paper was first to come up with ways to implement Deep Learning methods for the clinical text data and compare the results with more traditional Machine Learning classifiers. After vectorizing the text data using the top 100,000 most common words and limiting maximum features to 10,000 text labels, the results from our Deep Learning models are below. • Embedding Layer Results: Figs. 29.6 and 29.7 show the K-fold = 3 accuracy and log-loss of the embedding layer models with Panda’s get_dummies function and Keras’ to_categorical utility, respectively. Overall, Panda’s get_dummies function performed with 61.79% accuracy, which was 1–2% better than Keras’ to_categorical utility that we implemented for this paper. Figure 29.8 displays a sample graph of an Embedding Layer network of training and validation loss on the test data, whereas their accuracies are shown in Fig. 29.9. As seen from the large gap between training and validation accuracy on the test data in Fig. 29.9, we clearly had an overfitting issue to overcome. To develop an ideal model when tunning the hyperparameters on the validation text data, we modified the Embedding and Dense layers, experimented with different batch and epoch sizes, used other activation functions and optimizers, and applied regularization techniques like Dropout and SpatialDropout1D functions. As a
242
F. Kocaman et al.
Fig. 29.5 Architecture of the proposed technique for Machine Learning algorithms
Fig. 29.6 Results from Embedding Layer model using Panda’s get_dummies
Fig. 29.7 Results from Embedding Layer model using Keras’ to_categorical
result, the model presented in Fig. 29.2 gave the best results from iterated K-fold validation, as shown in Fig. 29.6. • Bidirectional LSTM Results: Since we could not solve the overfitting issue in the Embedding Layer model, we thought that our 3316-text data was the problem causing the models not to function well. Instead, we decided to experiment using Bidirectional LSTM because they are useful on Natural Processing problems and could work well with smaller text data [13]. Our Bidirectional LSTM model has an accuracy of 58.89% with a log-loss of 1.2205. Figures 29.10 and 29.11 exhibit the graph of log-loss and accuracy, respectively. Our Bidirectional LSTM model, like the Embedding Layer model, struggled with the overfitting problem. • BERT Text Augmentation Results: We realized that we needed more text data to mitigate the overfitting issue
from the two models above. Neural Network models perform better with having a large amount of data. The new text data augmented with BERT Contextual Embeddings from NLPAug library was used in the Embedding Layer. K-fold = 3 cross-validation was applied to evaluate this model. The K-fold combines the augmented train text data with the original test text data. Then the K-fold splits the entire input into training and testing sets at each fold. It trains on the training set and evaluates the test set every time. As Fig. 29.12 shows, the average accuracy was about 80.67%, with a log-loss of 0.6391, the highest accuracy we have achieved so far. Moreover, the weighted average precision, recall, and F1 scores were all 0.81. We also tested our augmented text embedding layer model only on the original test data, which gave us about 62.65% accuracy with a log-loss of 1.2061.
29 Performance Comparison Between Deep Learning and Machine Learning Models for Gene Mutation-Based Text…
Fig. 29.8 Embedding Layer test data los-loss graph
Fig. 29.10 Bidirectional LSTM test data los-loss graph
Fig. 29.9 Embedding Layer test accuracy graph
Fig. 29.11 Bidirectional LSTM test data accuracy graph
• Machine Learning Classifier Results: Our Random Forest classifier performed with an accuracy of 67.02%, the weighted average precision of 0.70, recall of 0.67, F1 score of 0.65, and the log loss of 1.0523, whereas the accuracy from the Stacking classifier was 66.27% with the weighted average precision of 0.68, recall of 0.66, F1 score of 0.66, and the log loss of 1.0344.
29.5
Conclusion and Future Work
Table 29.1 shows the comparison of the models we implemented in our work. Again, traditional Machine Learning algorithms prevailed over more complex Deep Learning methods. However, the Deep Learning algorithm with BERT augmented text data exhibited much better accuracy.
243
Future work involves using pre-trained word embedding techniques such as GloVe to capture semantic relationships within the text data, experimenting with other powerful Machine Learning algorithms like Extreme Gradient Boosting (XGBoost), and exploring other text augmentation techniques. Combining Machine Learning algorithms applied to the clinical text data with medical image processing techniques would further improve advancements in the diagnostic process and support all medical personnel. Moreover, when there is more study on classifying gene mutation for cancer in personalized medicine with the help of Machine Learning algorithms, it would also open doors to other areas of medicine such as curing diseases other than cancer and even early detection of an anomaly in infectious diseases like Covid-19. Acknowledgment This research was sponsored by the NATO Science for Peace and Security Programme under grant SPS MYP G5700.
244
F. Kocaman et al.
Fig. 29.12 Results from Embedding Layer augmented text model
Table 29.1 Classification test results Classifier Embedding Layer Embedding Layer Bidirectional LSTM Embedding Layer using Augmented Text Embedding Layer using Augmented Text Random Forest Stacking Classifier
Feature extraction Keras Tokenizer/panda’s get-dummies Keras Tokenizer/to_categorical Keras Tokenizer/panda’s get-dummies Keras Tokenizer/panda’s get-dummies/K-fold=3 Keras Tokenizer/panda’s get-dummies/Testing on Original Test data One-Hot encoding/TF-IDF Vectorizer One-Hot encoding/TF-IDF Vectorizer
References 1. National Cancer Institute (NCI), Cancer Statistics, September 25, 2020. Available online at https://www.cancer.gov/about-cancer/ understanding/statistics. Accessed 4 Nov 2021 2. S. Watson, Guide to Lung Cancer Mutations, July 16, 2020. Available online at: https://www.healthline.com/health/nsclc/guide-tolung-cancer-mutations. Accessed 4 Nov 2021 3. X.S. Zhang, D. Chen, Y. Zhu, C. Che, C. Sue, S. Zhao, X. Min, F. Wang, A Multi-View Ensemble Classification Model for Clinically Actionable Genetic Mutations. arXiv:1806.09737v2 [cs.LG], 17 Mar 2019 4. National Cancer Institute (NCI), The Genetics of Cancer, October 12, 2017. Available online at: https://www.cancer.gov/aboutcancer/causes-prevention/genetics. Accessed 4 Nov 2021 5. Memorial Sloan Kettering Cancer Center (MSKCC), MSKIMPACT: A Targeted Test for Mutations in Both Rare and Common Cancers, 2021. Available online at https://www.mskcc.org/mskimpact. Accessed 4 Nov 2021 6. Kaggle, Memorial Sloan Kettering Cancer Center (MSKCC) Personalized Medicine: Redefining Cancer Treatment, 2017. Available online at: https://www.kaggle.com/c/msk-redefining-cancertreatment. Accessed 4 Nov 2021 7. P. Singh, S.P. Singh, D.S. Singh, An introduction and review on machine learning applications in medicine and healthcare, in IEEE Conference on Information and Communication Technology (CICT) (2019) 8. S. Saxena, S.N. Prasad, Machine learning based sensitivity analysis for the applications in the prediction and detection of cancer
9.
10.
11.
12.
13. 14.
15.
16.
Test accuracy 61.79% 60.34% 58.89% 80.67% 62.65%
Test log-loss 1.2486 1.2795 1.2205 0.6391 1.2061
67.02% 66.27%
1.0523 1.0344
disease, in 2019 IEEE International Conference on Distributed Computing, VLSI, Electrical Circuits and Robotics (DISCOVER). https://doi.org/10.1109/DISCOVER47552.2019.9008083 (2019) G. Li, B. Yao, Classification of Genetic mutations for cancer treatment with machine learning approaches. Int J Design Anal. Tools Integr. Circ. Syst. 7(1) (2018) R.N. Waykole, A.D. Thakare, Intelligent Classification of Clinically Actionable Genetic Mutations Based on Clinical Evidences, 978-1-5386-5257-2/18/IEEE. (2018) P.S. Kohli, S. Arora, Application of machine learning in disease prediction, in 2018 4th International Conference on Computing Communication and Automation (ICCCA). https://doi.org/10.1109/ CCAA.2018.8777449 (2018) A. Singh, S.K. Jain, A personalized cancer diagnosis using machine learning models based on big data, in Proceedings of the Fourth International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC) IEEE Xplore Part Number:CFP20OSVART; ISBN: 978-1-7281-5464-0 (2020) F. Chollet, Deep Learning with Python. ISBN 9781617294433 (2018) R. Sangani, Powerful Text Augmentation Using NLPAUG, June 11, 2021. Available online at: https://towardsdatascience.com/ powerful-text-augmentation-using-nlpaug-5851099b4e97. Accessed 4 Nov 2021 Nithilaau, A Python Library to Augment Your Text Data, August 25, 2021. Available online at: https://www.analyticsvidhya.com/blog/ 2021/08/nlpaug-a-python-library-to-augment-your-text-data/. Accessed 4 Nov 2021 S. Marsland, Machine Learning an Algorithmic Perspective (2nd edn). Chapman & Hall/CRC Machine Learning & Pattern Recognition Series (2015)
Stock Backtesting Engine Using Pairs Trading
30
Rahul Chauhan, Marian Sorin Nistor, Doina Bein, Stefan Pickl, and Wolfgang Bein
Abstract
In this paper we present a Stock Backtesting Engine which would test historical data using pairs trading strategy. We implemented pairs trading strategy and ran it on historical data. We collect the S&P 500 data from the Internet and store it in a database. We then allow users to enter a set of stocks to find cointegrated pairs among them. We also provide an option to find the cointegrated pairs in all of S&P 500 stocks. Once the cointegrated stocks are selected we run the backtesting algorithm on these pairs and find from a given set of stocks, all the pairs of stocks that exhibit cointegration properties. Once such pairs are identified, this program would use pairs trading methods to calculate trades for each stock. Finally we provide the analysis of trades executed by the algorithm with average and daily data, and plot a chart of daily profit and loss with pairs trading strategy to showcase the effectiveness of the trading strategy. Keywords
Algorithmic trading · Automated trading · Investment bank · Pair trading · Stock market · Stock backtester · Stock correlation · S&P 500 stocks · Trading strategy · Heat map R. Chauhan · D. Bein () Department of Computer Science, California State University, Fullerton, Fullerton, CA, USA e-mail: [email protected]; [email protected] M. S. Nistor · S. Pickl Department of Computer Science, Universität der Bundeswehr München, Neubiberg, Germany e-mail: [email protected]; [email protected] W. Bein Department of Computer Science, University of Nevada, Las Vegas, Las Vegas, NV, USA e-mail: [email protected]
30.1
Introduction
Every year, investment banks spend billions of dollars to invest in technology with the objective of facilitating efficient flows of funds globally and providing world class services. For example, in 2019 JPMorgan, the highest spender, has a $11.4 billion technology budget this year, a 5.6% uptick from last year’s $10.8 billion. Bank of America’s IT spend was second, at $10 billion, followed by Wells Fargo at $9 billion and Citigroup at roughly $8 billion [1]. Besides investing in infrastructure, such as network security and large-scale database management, automated trading algorithms are also an increasingly mission-critical area of development. The internet has changed traditional stock market trades to electronic trades. Automated trading is now the most important innovation in the field. To have bots trade efficiently and effectively in stock market, they need to be able to test out their strategies and hypothesis in a safe environment. A stock backtester provides a perfect sandbox environment where these strategies can be tested out with historical data. It provides an important reference to sharpen the strategies before they can be applied in real world, where the risks are much higher. One of the most popular and simple strategy is of pairs trading [2, 3]. Pairs trading belongs to the statistical arbitrage class of trading strategies and is a simple trading strategy which involves finding a pair of stocks that exhibit similar historical price behavior, and then betting on the subsequent convergence of their prices in the event that they diverge [4, 5]. Since this trading style is very flexible and applicable to all asset classes, the methods to identify pairs trading signals is a widely researched topic. Pairs trading belongs to the statistical arbitrage class of trading strategies [6, 7]. In this paper we present a Stock Backtesting Engine which would test historical data using pairs trading strategy. This
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_30
245
246
R. Chauhan et al.
program would find from a given set of stocks, all the pairs of stocks that exhibit cointegration properties. Once such pairs are identified, this program would use pairs trading methods to calculate trades for each stock. Finally these trades would be analyzed and presented to user in the form of a simple report as well as visualization of different aspects of effectiveness of the trading strategy. The paper is organized as follows. In Sect. 30.2 we present background work, followed by our proposed system architecture in Sect. 30.3. Results for eight major companies for which we used pair trading are presented in Sect. 30.4. Concluding remarks and future work are given in Sect. 30.5.
30.2
Background Work
Algorithmic trading can be defined as having a machine systematically send orders to the trading exchange to be executed on behalf of human traders. Algorithmic trading uses computer codes and chart analysis to enter and exit trades according to set parameters such as price movements or volatility levels. Once the current market conditions match any predetermined criteria, trading algorithms can execute a buy or sell order on your behalf. Algorithmic trading allow automating a repetitive and non-value adding task in stock testing. This frees up humans in salespeople and traders to focus only on Selling and trading. Most importantly, having tools that can provide insight to trades to make informed decisions is a huge value added by technology. Furthermore, computers are far superior to perform such tasks which have low margins like arbitrage strategies to humans. Professional algorithmic traders often employ two major classes of algorithmic trading strategies, namely high frequency trading (HFT) and statistical arbitrage trading. HFT is characterized by high turnover rates and high order-to-trade ratios that leverage high-frequency financial data, usually in the scale of mili- to micro-seconds. In this paper our focus is mostly on the second method which is Statistical arbitrage trading. Statistical arbitrage involves simultaneously buying and selling assets that normally have the same price trends but whose prices diverge in the short term. This form of trading relies less on the speed of sending orders and more on the mean reverting nature of financial products, i.e., the phenomenon in which asset prices and returns eventually return back to the long-run mean of the historical time series. Pairs trading is a one of the statistical arbitrage class of algorithmic trading strategy. The rationale for it is that for similar companies or financial products, they should be priced similarly based on no-arbitrage pricing principle [8]. In other words, products with the same risk should have the same reward. Algorithmic Stock trading is quite popular nowadays to check the accuracy of the newly created strategy ormake sure
if the existing strategy works well on the given data stock backtesting engine also gains its popularity. Like the stock trading bot, the stock backtesting software comes in both paid and free software to help traders. The popular backtesting engine uses multiple trading strategies and, on its use, can have more than one option. Tradingview is a web-based platform for stock backtesting. Tradingview also provides to check the strategies into the real-time data similar to paper trading. Tradingview provides price chart data of given stock data, the total profit, trades performance on the charts, the percentage of profit, and provide advice to buy and hold. It is easily accessible because the web-based platform, accurate financial data, provides backtesting not only on stock but also for Forex and cryptocurrencies. The MetaTrader5, also known as MT5, with backtesting also provides expert advisors. The client must select the trading strategy before, and the corresponding advisor will get selected. The advisor program needs to run on historical data with some initial parameters. The advisor then runs a trading strategy several times with different settings on historical stock data. The process later allows the user to select a trading strategy from different combinations. MT5 Strategy backtesting engine works on most of the currency. MT5 is a program that needs to install on the user’s computer and generally installed by one of the MT5 agents on the user’s computer. MS Backtesting is another popular stock backtesting engine. It allows users to do 58 different backtesting. Once the backtest is complete, MS backtesting will show the use list of every buy or sell trade and also provide a portfolio chart. Interactive Brokers’ Portfolio Builder is created by Interactive Brokers, it is meant to be used for testing different strategy on historical data. This tool starts with creating an investment strategy, where we can choose different rules. It then allows you to select initial parameters like investment amount and long/short leverage. It is a powerful tool to visualize the investment strategy. Trend Spider backtester uses a different approach to backtesting. This backtesting engine is built in a way it can detect trend lines and Fibonacci patterns automatically. TrendSpider allows use to write code for trading strategy and run test it. Backtesting a trading algorithm means to run the algorithm against historical data and study its performance. A backtesting can provide confidence for any trading strategy because giving a positive result on previous data can also mean the same for future data. Backtesting can provide multiple data which can help to improve any strategy without using real money investment it can give for each trade will have an associated profit or loss and total of this profit/loss over the profit of the strategy. Some of major reason to perform backtesting are:
30 Stock Backtesting Engine Using Pairs Trading
– Filtration – At the beginning any trading strategy starts with generalization with backtesting one can remove all unnecessary conditions which are not useful in data and also can add any critical condition for any set of data. – Modeling – backtracking allows trading in any particular time frame to test specific conditions of any trading strategy. – Optimization – allowing to perform algo trading in large data can also give ideas for what can happen in future and if there is any way to optimize current conditions. – Verification – As mentioned multiple times in this documentation, without checking a new strategy performed in the real time market will be a disaster. If the technique is already performed in some data can give positive results and can give developers confidence to use it for future trades. – Net profit or loss – Net percent gained or lost for given stock. – Averages – Percentage average gain and average loss. – Ratios – Wins-to-loss ratio – Money estimation – In backtesting use can set desired amount backtesting will provide the result for only that given amount. Pairs trading is a market-neutral profits strategy that uses statistical and technical analysis. The pairs trading, as the name suggests, involves stocks in pairs. Pairs trading need two or more stocks with high correlation. Two companies with similar characteristics and work in a similar sector can have a high correlation between stocks. The strategy involves taking a long position in one stock and a short position in another stock. The following example can explain a pairs trading strategy. Assume there are two stocks X and Y which belong to similar market sectors and both financially identical in terms of revenue earned and expense structure. Both companies, X and Y, are in similar market sectors. Because they carry similar risks, so in theory, both should have similar prices. Although both companies X and Y carry a similar price due to fluctuations in supply and demand, the prices of X and Y may change, which leads to a temporary diverge in the price of X and Y. The pairs trading strategy takes advantage of this situation by selling the stock, which is relatively overpriced stock, and buying the other stock, doing similar when other stocks become overpriced and buy the underpriced stock. The pair trading strategy takes advantage of the mean-reverting behavior of the difference between the prices of the stocks X and Y. Other than market-neutral strategy, another reason why pairs trading is popular is that once a trader finds a pair, they only have to look for the relative price of a pair. This is much easier than trading with one stock with the help of a directional view. A real-world example of the correlation between the price trends of the stock are the stock trends of Pepsi and Coca Cola
247
Stocks (see Fig. 30.1). Both companies are very similar and their primary product is soda pop. Some of the other pairs can be Domino’s Pizza (DPZ) and Papa John’s Pizza (PZZA), Target Corporation (TGT) and WalMart (WMT). Cointegration tests analyze non-stationary time series, processes that have variances and means that vary over time. In other words, the method allows you to estimate the longrun parameters or equilibrium in systems with unit root variables [9]. Two sets of variables are cointegrated if a linear combination of two variables is lower than the order of integration. Following is the Definition of cointegration: xt and yt are said to be cointegrated if there exists a parameter α such that ut = yt − αxt is a stationary process. The cointegration exists between two sets of variables if a set of I(1) can create linear combinations with I(0). I(1) can give us information that a single set of differences can transform the non-stationary variables to stationery. Three main methods can perform cointegration test are: EngleGranger Two-Step Method, Johansen Test, and Maximum Eigenvalue test. In this paper we perform cointegration by the EngleGranger method. The Engle and Granger two-step method was developed by Robery Engle and Clive Granger in their seminal paper. One of the strengths of this procedure is the ease of testing. The two-step procedure is easy to follow and paints a good picture of the general idea behind cointegration. The idea is to first verify that the individual series are indeed integrated of order I(1) (being non-stationary). Then we regress one on the other using standard OLS and check if the residual series is integrated of order I(0) (suggesting stationarity). If the residuals are stationary, then we can extract the coefficients from the OLS model and use those to form a stationary linear combination of the two-time series. Suppose Xt and Yt are the natural logarithms of the price series of 2 assets, X and Y. The time series Xt and Yt is integrated order 1, and if there is a linear combination that exists Zt = Xt − βYt − α where Zt is a stationary time series. To check if time series Xt and Yt are cointegrated in Engle and Granger two-step test 1. If both Xt and Yt are non-stationary time series, first estimate the β and α using ordinary least squares (OLS) regression. 2. Then Zt is tested for stationarity where the AugmentedDickey-Fuller. If Zt is stationary, then we can say that Xt and Yt are cointegrated.
248
R. Chauhan et al.
Fig. 30.1 Stock trends of PEP and KO. (Image is reference from The Street.com)
There are a few shortcomings of the Engle-Granger procedure. It is incapable of simultaneously testing for cointegrating relationships among multiple series. In Engle and Granger method, the choice of which series to regress on the other is somewhat arbitrary. Depending on which variable we regress on we would receive a different cointegrating vector. Finally, Engle and Granger method overlooks the underlying error-correction model influencing the spread series. All of the above shortcomings can be addressed through the Johansen test. This procedure estimates cointegrated VARMA(p,q) in the VECM (vector error-correction model) form for m cointegrating relationships between k different series in xt . Δxt = αβ xt−1 +
p−1
Φi∗ Δxt−1 + at −
i=1
q−1
Θj at−j
1. Load configuration files into configuration objects. 2. Load data from the Internet to the database. 3. Find correlated pairs among the list of stocks using PairFinder Module. 4. Generate trade files and analysis using Backtest module
j=1
The notations, α and β are both k x m matrices, Δxt represents the first difference as x = x − x−1 , i are the AR coefficients, and j are MA coefficients. The cointegrating equation is defined by β x{t − 1}, where β contains the coefficients for the m cointegrating vectors.
30.3
the backtesting algorithm on these pairs. Finally, we provide the analysis of trades executed by the algorithm with average and daily data, and plot a chart of daily profit and loss with pairs trading strategy. The general architecture is shown in Fig. 30.2. The Main Module acts as a broker between different modules and is also a configuration loader. It loads the configuration files in their objects and passes these configuration objects to different modules. The primary functions of the main module are:
Our Proposed System
We implemented pairs trading strategy and ran it on historical data. We collect the S&P 500 data from the Internet and store it in a database. We then allow users to enter a set of stocks to find cointegrated pairs among them. We also provide an option to find the cointegrated pairs in all of S&P 500 stocks. Once the cointegrated stocks are selected we run
The database DB Module provides a repository for the entire stock data. The stock data is used in both finding the cointegrated pairs as well as calculating trades from the daily trade data for backtesting. The DB modules load the stock data and ticker from the Internet using Wikipedia and Yahoo finance. For efficiency it uses a local database for cache to store the trading data. The primary functions of DB modules are: 1. Create local database and tables. 2. Get all the S&P 500 tickers from Wikipedia. 3. Load daily stock data for these tickers from Yahoo finance into the local database tables. 4. Provide functions to other modules for calculating data from local database.
30 Stock Backtesting Engine Using Pairs Trading
Configuratuion Files
249
Pair Finder
Find cointegrated pairs
Get DB Config
Get Backtest Config
Main
Run backtest on cointegrated pairs
DB
BackTest engine
Stock database (Postgresql)
Stock data from Yahoo Finance
Analyze Trade data
Analysis Module
Trade Files Get trade data
Fig. 30.2 Detailed design diagram for the system
Pair Finding Module uses stats models tools to find correlation between stock data time series. The correlation is measured as pValue. Once the pValues are calculated we measure it against a pValue threshold and decide whether two stocks are cointegrated. This module also displays a heat map of the pValues of multiple stocks, which gives a nice visualization of which stocks are correlated. The primary functions of this module are: 1. Get all the data for user given Stock tickers from the db module. 2. Calculate pValues for each stock pair. 3. Create a heat map for every sector of stocks. 4. Return correlated stock pairs. Backtest Module takes the correlated pairs and generates trade files using the pairs trading algorithm. The central tenet of this algorithm are zValues. The current zValues for each pair are calculated using short and long look up windows and standard deviation. These zValues are then compared with the threshold values set in configuration. Based on these the backtest module produces trade signals and writes trades in Trade files. The primary functions of Backtest Module are: 1. Calculate zValues for each pair.
2. Using zValues and thresholds create trade signals and make buy/sell decisions. 3. Write these decisions into Trade files for persistent storage. 4. Call Analysis module to use Trade files and generate summary of trade data. Analysis Module collects all the trade information in the Trade files and then generates a summary of all the trades executed. It provides an average trade summary as well an average daily trade summary. It also generates a chart for daily profit and loss with Pairs trading strategy. The primary functions of Analysis modules are: 1. Collect trade data from trading files. 2. Calculate overall and daily profit/loss and averages for trades. 3. Plot daily profit and loss on chart. The configuration for database connection and backtesting engine is stored using YAML files for easy readability. This also allows us to modify program config without recompiling it. We use PyYaml for parsing the config files into data classes and pass the objects of these data classes to other modules. To find pairs we use the StatsModel python library. This provides a function to calculate pValue between two given
250
time series. The higher the pValue the less correlated are the two time series. We compare these pValues to a configurable threshold to make sure we find two correlated pairs. Backtesting module takes the list of pairs and for each pair calculate their zValues. The zValues are calculated by finding ratios between the two stocks of the pair and doing a rolling average and standard deviation calculation at each data point (day). It uses stock data ratios and rolling mean and standard deviation to calculate the zscore. Once we have the zValues for Stocks at every data point, we compare the value with configurable upper and lower limits of zValues and check what kind of trade we are in right now. If we are in a short trade and the logic dictates that we should continue to be in a short trade we do not create another short position. Similarly for a long trade we do not want to write in trade files multiple long trades. To make sure that it does not create extra trades and reduce the effectiveness of our trading strategy, we must maintain the state to store our current trade. Which means if we are in a short trade and get a short signal we ignore it, similarly if we are in a long trade and get a long signal, we ignore it as well. Trading signals are only used to either enter or exit from a trade. So if we are in a particular trade and get a signal for its reversal or lack of effectiveness we exit from such trade. After the trades are calculated and trade files are generated, the analysis module just collects all the trade data and computes daily and overall average, profit and loss, winning trades etc. to display the summary of success of strategy to the user. This also generates a chart for the daily profit and loss. We use Python’s matplotlib library to render the charts. We also compute some moving averages to show on the chart, which makes it easy to track the profit/loss progress on the chart. We also plot moving averages on this chart. Moving averages are defined by average of different subsets of a given data set. These subsets are determined by the window size. In finance it is usually a price indicator that can be used for technical analysis. The primary reason for plotting a moving average on a chart is to smooth out a big fluctuation. This is precisely why our program plot Moving averages in this chart. We want to be able to see the long-term price action of our trades, and moving averages help us to ignore the daily noise and smooth out the graph to see the effectiveness of our trade. We plot two different moving averages in this chart, one with back looking window of 50 data points, while the other is for 200 data points. Each data point in this graph refers to a day. These two Moving averages provide a clear and visual identification of the effectiveness of our pairs trading strategy.
R. Chauhan et al.
30.4
Results Based on Yahho Finance Data
We use the Stats Model tools to find cointegrated pairs of stocks. This is done by generating pValues of the stocks. These values are displayed in a heat map for each stock of the same sector. In Fig. 30.3 we show the heatmap for correlation pValue score between following tech sector stocks {‘GOOG’, ‘FB’, ‘TWTR’, ‘AMZN’} representing Google, Facebook, Twitter, and Amazon. Different colors represent different level of correlation between stocks. We note that, lower the value of pValue between the stock pairs, the more correlated they are, and make good candidates for pairs trading strategy. In heatmap the higher to lower correlation is shown with a green to red spectrum. As we can see in the above example, two stock pairs, GOOG and AMZN are most correlated, while GOOG and FB are only slightly less correlated. Both pairs are good candidates for pairs trading strategy. On the other hand, TWTR stock is not correlated with any other ticker, which is shown by all red heat zones in TWTR row. Our system also displays the price data for these stocks to provide a visual confirmation of a good correlation (see Fig. 30.4). We note that GOOG, AMZN and FB, GOOG pair really do correlate, and this confirms the correctness of heatmap generated. A similar example is shown for the retail sector {‘WMT’, ‘TGT’, ‘LOW’, ‘HD’} representing Walmart, Target, Lowe’s and Home Depot. Figure 30.5 the heatmap for correlation pValue score between following four retail sector stocks
Fig. 30.3 Heatmap of correlation between four tech sector stocks
30 Stock Backtesting Engine Using Pairs Trading
251
Fig. 30.4 Price data for the four tech sector stocks
{‘WMT, ‘TGT’, ‘LOW’, ‘HD’}. Different colors represent different level of correlation between stocks. The price data is shown in Fig. 30.6. Both price data as well as heat map, LOW and HD are the only highly correlated stocks in this set of stocks. The system creates a trade summary from analyzing the master results and pair files. The daily profit and loss chart is also plotted (see Fig. 30.7). This is a daily profit / loss chart that shows at every data point (day) how much profit or loss our trading strategy has yielded. The daily profit and loss chart shows how effective our trading strategy is on a day-to-day basis. It is important to realize at this step is that no trading strategy could produce winning trades every datapoint or even guarantees winning trades as a higher percentage of total trades. Which is why we also generate a report that shows how many winning trades were performed by our strategy. This is only the indication of total number of winning trades and not whether the strategy was successful or not. This is because it is possible to wipe out a lot of winning trades with a single losing trade and so while the winning trades would be a higher percentage of all trades, the total initial capital might still be reduced.
Fig. 30.5 Heatmap of correlation between four retail sector stocks
252
R. Chauhan et al.
Fig. 30.6 Price data for four retail sector stocks
30.5
Conclusion and Future Work
This system has some limitations. Our system program does not allow for plug in strategy backtesting since our backtesting algorithm is currently hard coded. We currently do not support any machine learning or deep learning for making future predictions. This program is strictly for backtesting the pairs trading hypothesis. It takes a significant amount of time to calculate suitable correlated pairs out of a large set of Stocks. We have the option to generate the correlated pairs out of all of S&P 500 stocks, indexed by sectors in the pair generator module. However, running that takes a very long time and so its much more convenient to provide a smaller set of stocks, from which the pairs can be computed quickly. We currently write the files with trading information. This could be improved by natively writing the files to a cloud service like S3. We do not have the option to go back and look at heat
map charts again. Adding an option to persist the image files generated during the running of system is desirable. Adding the forward trading to the existing project will become the full algorithmic trading bot. Forward trading is similar to back trading. Both forward/back trading use the same algorithm and logic. Cointegration can be improved by adding a Bayesian Kalman filter. A Bayesian Kalman filter could help improve the trading results by cutting down statistical noise and make more accurate predictions of future state of the service from historical data. We could convert the system into a web service that generates the trading data and send it back with REST APIs. This data could then be captured in a Frontend web framework like React and plotted on the browser. This would make it possible for the program to run on a more powerful backend fixing the limitation of taking a long time and we can still access it from any computer connected to it. We could train a model using different machine learning algorithms to make accurate predictions
30 Stock Backtesting Engine Using Pairs Trading
253
Fig. 30.7 Daily profit and loss chart
and then reinforce it by the results of these predictions. The idea of reinforced learning would make the trading bot much more accurate and increase trust in the system. Currently our backtesting engine only uses Pairs trading strategy. This pairs trading algorithm is also hardcoded. We could easily make the Engine separate from the actual trading algorithm so its easy to put different trading strategies to test and compare their data.
2.
3. 4. 5.
Acknowledgments This research was sponsored by the NATO Science for Peace and Security Programme under grant SPS MYP G5700.
6. 7.
References 8. 1. R. Shelvin, How Much Do Banks Spend On Technology? (Hint: It Would Weigh 670 Tons In $100 Bills), (2019, Apr 1). From Forbes.com https://www.forbes.com/sites/ronshevlin/2019/04/
9.
01/how-much-do-banks-spend-on-technology-hint-chase-spendsmore-than-all-credit-unions-combined/?sh=1066d949683a https://en.wikipedia.org/wiki/Pairs_trade%23: :text=A%20simpli fied%20example,-Pepsi%20(PEP)%20and&text=If%20the%20 price%20of%20Coca,to%20their%20historical%20balance %20point. (n.d.). From Wikipedia https://www.fidelity.com/learning-center/trading-investing/trading/ pairs-trading. (n.d.). From Fidelity https://www.investopedia.com/articles/trading/04/090804.asp. (n.d.). From Investopedia.com https://www.investopedia.com/terms/p/pairstrade.asp. (n.d.). From Investopedia.com https://www.statsmodels.org/stable/generated/ statsmodels.tsa.stattools.coint.html. (n.d.). From Statsmodels.org J.T. Peng, Research of the Matlab application in the fingerprint identification system, in International Conference on Image Analysis and Signal Processing, Hangzhou, (2012), pp. 1–5 S.H. Penman, Accounting for Value (Columbia Business School Publishing, 2011) B. Rao, Cointegration: For the Applied Economist, 2nd edn. (Springer, 2007), p. 279
Classifying Sincerity Using Machine Learning
31
Rachana Chittari, Marian Sorin Nistor, Doina Bein, Stefan Pickl, and Abhishek Verma
Abstract
Quora is an online platform that empowers people to learn from each other. On Quora, users can post questions and connect with others who contribute unique insights and quality answers. But as with any other social media or online platform, there is the potential for misuse. A key challenge in maintaining the integrity of such an online platform is to classify and flag negative content. On Quora, the challenge is to identify insincere questions. Insincere questions could be those founded upon false premises, are disparaging, inflammatory, intended to arouse discrimination in any form, or intend to make a statement rather than look for helpful answers. We propose to develop a text classification model that correctly labels questions as sincere or insincere on the Quora platform. For this purpose, we used the Quora Insincere Questions Classification dataset, which is available on Kaggle. We first trained classical machine learning models such as Logistic regression and SVMs to establish a baseline on the performance. However, to leverage the large dataset, we used neural network-based models. We trained several models including standard neural networks, and LSTM based models. The best model that we obtained is a two-layer Bidirectional LSTM network that takes word R. Chittari · D. Bein () Department of Computer Science, California State University, Fullerton, Fullerton, CA, USA e-mail: [email protected]; [email protected] M. S. Nistor · S. Pickl Department of Computer Science, Universität der Bundeswehr München, Neubiberg, Germany e-mail: [email protected]; [email protected] A. Verma Department of Computer Science, California State University, Northridge, Northridge, CA, USA e-mail: [email protected]
embeddings as inputs. The classification accuracy and F1score for this model were 96% and 0.64, respectively. Keywords
Bi-directional LSTM model · Insincere questions · Logistic regression · Machine learning · ML · Negative context · Text classification · Quora dataset · LSTM model · SVM model
31.1
Introduction
In addition to the search engines such as Google and Bing, there are websites known as question forums that one can use to gain knowledge. A question forum is an online discussion site where people can hold discussions in the form of posted messages. They have gained a lot of popularity due to their easy to use and understand methodology. Quora, Stack Overflow, Yahoo Answers are some examples of question forum websites. Quora is a popular online platform for users to ask simple, personal, professional questions and generally obtain well thought-out answers to [1]. While most of the questions are asked in good faith, there are several instances where people tend to ask questions that are inappropriate or inflammatory. They may be targeted at a specific group of people, intend to make a false statement, or sometimes make no sense. These questions tend to create havoc and threaten the integrity of the platform. Questions like these are termed “insincere” and they deviate from the main purpose of an online forum, which is simply to help users share knowledge. To ensure the safety of the users and the integrity of the forum, it is thus extremely important to classify, flag and remove such insincere questions before they can cause any significant damage. In the past, Quora made use of human
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_31
255
256
R. Chittari et al.
reviewers to label questions as sincere or insincere. However, due to the tremendous growth in the number of users and the number of questions posted daily, it becomes intractable to maintain a human based review system. Thus, it becomes important to develop an automated system to perform this classification task. One viable approach is to use a machine learning based text classification approach. We propose to build a supervised machine learning model that can classify questions as sincere or insincere. The paper is organized as follows. In Sect. 31.2 we present background work, followed by our proposed system architecture in Sect. 31.3. Results from applying various machine learning algorithm are presented in Sect. 31.4. Concluding remarks and future work are given in Sect. 31.5.
31.2
Background Work
Detecting inappropriate and negative content online is a highly relevant problem today. This is relevant not just to Quora, but to all social media platforms and online forums. Examples of classifying negative content include analyzing movie reviews on IMDB, entries on Wikipedia, and tweets on Twitter. These are essentially problems in the domain of natural language processing. A lot of work has been done using machine learning and deep learning models. The basic idea is that supervised machine learning algorithms can be trained on a labeled set of content to correctly identify negative examples. To this end, several companies have shared large pools of data on platforms such as Kaggle [2], and challenge machine learning researchers to develop state of the art classification models. In the past, traditional machine learning approaches were used to tackle these classification problems. The important consideration with these approaches is how to represent a sequence of words. Once a good set of features have been extracted, they can be used with almost any classification model. Typically, representations such as bigrams, n-grams, and bag-of-words, have been combined with classifiers such as Logistic Regression and Support Vector Machines. A few examples of these are the following. Vectorization using bagof-words was combined with Logistic Regression and Naïve Bayes classifiers for tweet classification in [3, 4]. A Support Vector Machine model was used to classify BBC documents into five categories in [5], and n-grams were combined with SVMs for emergency event detection on social media in [6]. Deep learning [7] is a broad family of machine learning models that use artificial neural networks [8, 9]. The term “deep” refers to the use of multiple layers of neurons in the network. The availability of massive data sets and high-speed computational resources such as Graphics Processing Units have enabled us to train deep neural networks to perform a wide variety of machine learning tasks.
Deep learning models have achieved near human performance in various tasks such as object recognition and speech recognition. They also provide the best results for natural language processing tasks such as machine translation, text classification and sentiment analysis. One of the important advantages of using neural networkbased models for natural language processing is that they can be used to learn word embeddings. A word embedding is a mapping from a word to a real-valued vector that is thought to encode the meaning of the word. While several methods exist to generate such embeddings, deep learning methods have provided some of the most widely used word embeddings, such as GloVe (Global Vectors for Word Representation) [10]. These word embeddings can be thought of a representational layer in a deep learning architecture. Several types of deep learning architectures have been considered for text classification. These include convolutional neural network models, sequence models such as recurrent neural networks, and attention-based models [11] provides a comprehensive review of more than 150 deep learning models for text classification that were developed in the last few years. In this project, I have mainly focused on using recurrent neural networks and their variants for insincere question classification. We used an open dataset available on Kaggle [2]. There are about 1.3 million examples in this dataset. Each example consists of 3 fields: q_id, question_text and target: q_id: Unique question ID assigned to each question question_text: a question posted on Quora in the English language target: takes the values 0 or 1. The values 0 and 1 correspond to sincere and insincere questions, respectively. Here are a couple of example questions and their target values: Sincere (0): “How did Quebec nationalists see their province as a nation in the 1960s?” Insincere (1): “Which babies are more sweeter to their parents? Dark skin babies or light skin babies?” Many of the insincere questions are discriminatory in nature, inflammatory, politically polarizing, and potentially harmful to raise on the Quora platform. It is thus crucial to correctly identify and weed out these questions. We analyzed was the distribution of the target values. We found that the Kaggle dataset is highly imbalanced. Only about 6.2 % of all the examples are labeled as insincere. This implies that we should not use classification accuracy as a metric to evaluate our model. For instance, a naïve model that always outputs 0 as its prediction would already achieve a classification accuracy of 93.8%, but importantly it would never be able to correctly identify an insincere question (with target value 1). Therefore, it is important to use metrics such
31 Classifying Sincerity Using Machine Learning
as precision and recall to correctly evaluate the performance of a model on this dataset. We applied the following preprocessing steps: – Convert all text into lowercase – Remove all special characters and non-English characters that appear in the dataset – Remove punctuation characters – Remove stop words or common English filler words like “and”, “an”, “the” – Split the question strings into word tokens We utilized the Natural Language Toolkit (NTLK) [12] to perform several of these preprocessing steps. NTLK is an open-source platform for building Python programs to work with natural language data. It provides a corpus of English stop words and tokenization functions that I used to clean my dataset. Alternatively, the Torchtext library of PyTorch can also perform the same set of preprocessing steps. PyTorch is an open-source machine learning framework that can be used to quickly build deep learning models. The next stage was to extract features from the cleaned dataset that can be used to train the machine learning models. For text classification, we must first convert the text into a vector representation. Word vectorization is the process of converting text into a numerical representation and is an important first step in any Natural Language Processing task. We have considered the following two approaches for feature extraction: (a) Word count vectors. A word count vector of a question or sentence is a representation that specifies how many times each word in a given vocabulary appears in the sentence. This is essentially a bag-of-words representation, which disregards word order but keeps multiplicity. To implement the word count representation, we first built a dictionary of all the unique words that appear in my dataset. Each question was then converted into a vector of word counts. The length of this vector is equal to the number of words in the dictionary. Consider the following example for illustration. Suppose that a dictionary comprises the following words: {“hello”, “who”, “how”, “are”, “you”, “me”, “doing”, “today”}. Let us see what the word count vector for a sample question, “Hello, hello, how are you?” would like. Since the dictionary has 8 words, the sample question is represented by an 8-dimensional vector, with each element corresponding to the count of a particular word in the dictionary. The word count vector for our example would be: (2, 0, 1, 1, 1, 0, 0, 0). While word count vectors are useful in some cases, they do have a few limitations. When the vocabulary is large, the word count vectors are sparse and high dimensional. The Quora dataset, for instance,
257
has about 100,000 unique words. And each question in the dataset typically has only 10–20 words. Further, the word count vectors ignore word order, which is certainly important for assessing the nature of a question. (b) Word embeddings are mappings from words to Ndimensional real-valued vectors such that any two words that are closer in the vector space are similar in meaning [13]. Various methods can be used to generate these mappings, including deep learning techniques. Further, in deep learning models, the embeddings can be thought of a representational layer in the architecture. There are several pre-trained word embeddings that are now available. A few examples are wiki-news-300d-1M, GoogleNews-vectorsnegative300, paragram_300_sl999, glove.6B.100D, and glove.840B.300D. For this project we have used the GloVe embeddings. GloVe or Global Vectors for Word Representation [10] is an unsupervised learning algorithm for obtaining vector representations for words. The glove.840B.300D embeddings output a 300dimensional vector for each word in the vocabulary. To combine the embedding vectors for words in a question, I considered 3 methods: 1. Computing the mean of embeddings vectors of all words in a sentence 2. Concatenation of the embedding vectors of all the words in a sentence 3. Use these embeddings as inputs to the sequence models The first approach is obviously very lossy. In the second approach, the dimensionality of the feature vector increases with the number of words in the question. We explored using the concatenated features as inputs to a few basic classifiers.
31.3
Machine Learning Models Applied
To obtain a baseline on the performance we started with Naïve Bayes Classifier, Logistic Regression, and Support Vector Machines. A Naïve Bayes classifier is a probabilistic classifier based on Bayes theorem, which makes the strong assumption that the features are independent of each other [14]. Given an input feature vector x = (x1 , x2 , . . . , xN ), we must compute the probability of every class given x. A vector x is assigned to the class Ci which has the highest probability. Logistic regression is a binary classification algorithm, used to model the probability of a certain class given a set of input features. In logistic regression, the probability of the target given input features is computed as follows: P (y = 1|x) = σ wT x + b
258
R. Chittari et al.
where w is a weight vector, b is a bias term and σ is the sigmoid function given by σ (z) =
1 (1 + e−z )
Let us call the prediction yˆ = P (y = 1|x). We can then compute the cross-entropy loss function for one example as follows: L y, yˆ = −y log yˆ − (1 − y) log 1 − yˆ The weight vector and bias term are then learnt such that they minimize the average cross-entropy loss over all the training examples. This optimization is typically performed using gradient descent-based methods. For the logistic regression classifier, we used a concatenation of 300-dimensional Glove word embedding vectors as inputs. We first computed the histogram of the number of words in each question for the entire dataset. Based on this histogram, I chose the maximum number of words, Nmax, to represent each question. Then, each question was represented by a feature vector of size 300 Nmax . For questions with fewer words than Nmax , we zero padded the feature vector to ensure they are of the same length for all examples. A Support Vector Machine (SVM) is a binary classification algorithm that tries to find the separating hyperplane/decision boundary with the maximum margin [15]. The SVM is a linear classifier, but they can also perform nonlinear classification by implicitly mapping the input features into higher-dimensional feature spaces. This is known as the kernel trick. We applied SVM for question classification with a concatenation of word embedding vectors as input features. We implemented the Naïve Bayes, Logistic Regression, and SVM classifiers using the scikit-learn, which is an opensource machine learning library in Python. One of the crucial limitations of the methods described previously is that they do not explicitly consider the sequence of words in a question. A family of artificial neural networks known as Recurrent Neural Networks (RNNs) can handle sequential inputs and are perfectly suited for the question classification task. A Recurrent Neural Network (RNN) is a class of neural networks with self or recurrent connections between nodes. They are derived from feed-forward neural networks, which have only forward connections between nodes in different layers of a network. The hidden layer in RNNs can be thought of as memory states, which are continuously updated by the inputs. Further, RNNs can be used to handle input sequences of variable lengths. This makes RNNs ideally suited for handling applications with sequential inputs such as machine translation, time-series prediction, speech recognition, and indeed text classification. Next, let us look at how a RNN can be trained.
Backpropagation is a widely used algorithm for training neural networks [16]. In supervised learning, training a model requires us to update its weights or parameters such that they minimize some loss criterion. This requires us to compute the gradient of the loss function with respect to all the model parameters. Backpropagation is an algorithm for computing the gradient of the loss function with respect to the weights of a feedforward neural network for a single input-output example. In contrast to a naïve direct computation of the gradient with respect to each individual weight in the network, backpropagation uses the chain rule of calculus to compute the gradients one layer at a time. It iterates backwards from the output layer to avoid redundant computations of intermediate terms. This results in a highly efficient algorithm for computing gradients in a neural network. This efficiency makes it feasible to use gradient methods for training multilayer neural networks. The backpropagation algorithm was originally designed for feedforward neural networks. But they can be adapted to train recurrent neural network as well. A recurrent neural network can essentially be unfolded in time to obtain a feedforward network with tied weights. This intuition was used to develop the Backpropagation Through Time algorithm (BPTT) [17]. Given an input sequence, BPTT works by unrolling all input timesteps. Each timestep of the unrolled recurrent neural network can be considered as an additional hidden layer, which also receives as input the hidden state from the previous time step. BPTT calculates and accumulates errors across each time step to compute the gradient of the loss function with respect to the network’s weights. BPTT is a useful algorithm for training recurrent neural networks. However, it can become computationally expensive as the number of time steps in the input sequence increases. Further, when the input sequences are long, the number of derivatives required for a single weight update will be high. This can cause the gradients of the weights to become diminishingly small depending on the nature of the activation functions used in the network. This problem is known as the vanishing gradients problem. For some activation functions, we can also observe the exploding gradient problem, which would result in numerical overflow problems. The vanishing gradient problem is a major drawback of the BPTT algorithm especially for standard recurrent neural networks that use sigmoidal activation functions. The vanishing gradients make learning extremely slow. Another consequence of the vanishing gradient problem is that is does not allow learning of long-range dependencies within input sequences. Learning dependencies among different parts of a sequence is vital for several tasks such as machine translation and text classification. Variations of BPTT such as Truncated BPTT were developed to solve the vanishing gradient problem. But perhaps
31 Classifying Sincerity Using Machine Learning
259
the best solution was networks that used gating mechanisms, such as Long Short-Term Memory networks [18]. Long Short-Term Memory (LSTM) is a type of recurrent neural network architecture that uses gating mechanisms to regulate the flow of information through a neuron [19]. A basic LSTM unit is composed of a memory cell c, hidden state h, and input, output, and forget gates that allow the unit to remember values over arbitrary times by regulating the flow of information in and out of the unit. In theory, standard RNNs could keep track of arbitrarily long-term dependencies in the input sequences. However, the problem with standard RNNs is computational in nature. While training RNNs with BPTT, we run into the issue of vanishing gradients. RNNs using LSTM can alleviate this problem because LSTMs can remember values across several time-steps if necessary. This would allow the gradient to flow unchanged during backpropagation. This makes LSTM based networks well-suited to problems involving data where there can be lags of unknown duration between important events in the input sequence. There are other types of architectures that use gating mechanisms such as Gated Recurrent Units (GRU) [20] that are based on the same intuition. Recurrent neural networks are ideally suited for processing sequential inputs. However, in normal RNN architectures the computation at any time step depends only on the inputs up to the current time. However, for several tasks it might also be important to consider input data in future time steps as well. This is crucial in applications such as natural language processing. In many cases, to understand the context in which a word is used, it is important to know what comes before and after it in the sentence. Consider the following two sentences:
very similarly to unidirectional RNNs, because the forward and backward states do not directly interact with each other. Bidirectional RNNs are especially useful when the context of an input is needed like in the Quora insincere question classification task. We decided to use a bidirectional LSTM (Fig. 31.2) for this classification task. This combines the advantages of both a bidirectional architecture and LSTM units. At each time step, the embedding vector for a word in the question text is provided as input to the network. The embeddings can be considered as an additional representational layer in the architecture. These embeddings can be learnt while training the network. Pre-trained embeddings can also be used (i) by keeping them fixed, or (ii) by finetuning them during training. The embedding vectors at each time step act as input to the forward and backward hidden states. The hidden states are usually initialized using random values. After the entire sequence has been processed in both directions, the final hidden states in both forward and backward directions are fed as inputs to an output neuron. The activity of this neuron is then passed through a sigmoid function to obtain the probability of the given question being insincere. Fig. 31.2 illustrates an architecture with one hidden layer. But this could be generalized to having multiple hidden layers. RNNs with multiple hidden layers are known as stacked RNNs. In a stacked RNN architecture, the hidden states in a given layer are fed as inputs to the hidden layer above it. Increasing the number of layers or depth of the network enhances its representational power, similar to conventional feedforward networks. It is also thought that stacking allows the hidden states at each level to operate at different timescales, and thus allow the network to learn both shortrange and long-range dependencies in a sequence [22].
(a) She said, “Teddy bears are for sale” (b) She said, “Teddy Roosevelt was the 26th president of the United States”
31.4
A recurrent neural network processing the two sentences would perform the exact same computations for the first three words. However, in the first sentence the word “Teddy” refers to a toy, and in the second it refers to a person. This information can be obtained only by using the words that follow “Teddy” in the two sentences. Bidirectional RNNs were developed to address exactly this issue. A bidirectional RNN is simply an RNN with two groups of neurons in the hidden layer. One group serves to process the input sequence in the forward direction, and the other for the backward direction (see Fig. 31.1). The two groups of neurons are then connected to the same output neurons, thereby conveying relevant information from both directions of processing. Bidirectional RNNs can be trained
We used Google Colaboratory, or “Colab” or short, to develop and train the bidirectional LSTM model. Colab is a product from Google research that allows one to write and execute Python code through the browser. Colab also allows one to use Graphics Processing Units (GPUs). GPUs can be used to train deep learning models much faster than on CPUs. Since the Quora Insincere Questions dataset is highly imbalanced, the classification accuracy is not a good metric to evaluate a model’s performance. A confusion matrix is a better way of summarizing the results. We can use the entries of the confusion matrix to compute three metrics: precision, recall, and F1-score. For our classification task, it is very important to correctly flag all the insincere questions. Therefore, recall is probably the more important metric to
Results of Various Machine Learning Algorithms
260
R. Chittari et al.
Fig. 31.1 Comparison of a unidirectional vs. bidirectional RNN [21]
Fig. 31.2 Bidirectional LSTM network architecture
consider. When we need a balance between precision and recall, we can use the F1-score. We used a subset of the full data to train the classical models. The full dataset has roughly 1.3 million examples. Using word count vectors or concatenation of word embeddings results in very high dimensional features. For instance, even for a subset of 200,000 examples, the size of the input matrix of word count vectors for all examples was 200,000 × 85,903. The number of unique words in the dataset is 85,903. As the number of examples used increases, we can expect the number of unique words to grow as well. Such high dimensional features result in memory constraints. Thus, to obtain a feasible implementation as well as a rough lower bound on the performance, I trained the models using small subsets of data. We used a subset of size 50,000 for Naïve Bayes, and size 20,000 for Logistic Regression and SVM. In each case, we split the data into train and test with a split ratio of 0.8, i.e., 80% of the data was used for training and the remaining 20% of the data was used for evaluating the model performance. Just for the Naïve Bayes classifier, we managed to train the model using a subset of size 200,000. We implemented the Bidirectional LSTM using PyTorch on Colab. PyTorch is an open-source machine learning plat-
form for building and training deep neural network models. One of the most powerful features of PyTorch is Autograd, which is an automatic differentiation engine that powers neural network training. In PyTorch, we just need to implement the forward pass of the neural network model and provide the loss criterion. The Autograd engine can automatically compute the gradients of the loss function with respect to all the weights in the network. This makes the training of a neural network using backpropagation almost trivial. The torchtext package in PyTorch consists of data processing utilities for natural language. We used torchtext to preprocess my data and created a PyTorch dataloader. We used the GloVe.840.300D embeddings to map input words to vectors. This is treated as an embedding layer in the neural network architecture. Our implementation also has the option to fine-tune the embeddings if required. We used cross-entropy as the loss function with the Adam optimizer for training my model. Adam [23], derived from Adaptive Moment Estimation, is a gradient-based optimization algorithm that computes adaptive learning rates for each model parameter. Finally, for training my neural network model, we split the dataset into training and validation sets with a split ratio of 0.9.
31 Classifying Sincerity Using Machine Learning
261
network with two hidden layers. We obtained better recall and F1 score than the baseline set by the traditional ML methods. The model most likely benefits from the advantages of the bidirectional architecture and the LSTM. This could be analyzed by comparing the performance of our model with that of a unidirectional architecture. While the performance of my model is quite good, there are several things that could be done to potentially improve the classification performance. The first of these is to perform a precision-recall analysis to find the best threshold. In classification problems such as these, there is always a trade-off between precision and recall. The precision-recall analysis would allow us to pick a threshold based on the criteria we want to satisfy. Fig. 31.3 Cross entropy loss with 64 hidden units Acknowledgment This research was sponsored by the NATO Science for Peace and Security Programme under grant SPS MYP G5700.
References
Fig. 31.4 F1-scores with 64 hidden units
We started with 500,000 examples for training the model. We used a 2 hidden layer architecture throughout. For the first pass, we used 64 units in each hidden layer, and the default learning rate of 0.1 for the Adam optimizer. Figures 31.3 and 31.4 illustrate the cross-entropy loss and F1-scores vs. no. of training epochs. One epoch corresponds to iterating through all the mini-batches of the dataset. In Fig. 31.3, although the mean training loss (blue) was continuously decreasing with the number of epochs, the validation loss increased after 4 or 5 epochs. This suggested that the model was in the overfitting regime. We obtained similar overfitting effects for different values of the number of hidden units. In order to address overfitting, we plan to use Dropout regularization [24].
31.5
Conclusion and Future Work
The neural network model that we used for the Quora Insincere question classification task was a Bidirectional LSTM
1. Quora, [Online]. Available: https://www.quora.com/ 2. Kaggle, [Online]. Available: https://www.kaggle.com/ 3. O. Aborisade, M. Anwar, Classification for authorship of tweets by comparing logistic regression and naive bayes classifiers, in IEEE International Conference on Information Reuse and Integration (IRI) (2018). https://www.fidelity.com/learning-center/ trading-investing/trading/pairs-trading (n.d.). From Fidelity 4. S.T. Indra, L. Wikarsa, R. Turang, Using logistic regression method to classify tweets into the selected topics, in International Conference on Advanced Computer Science and Information Systems (ICACSIS) (2016) 5. A.W. Haryanto, E.K. Mawardi, Influence of word normalization and chi-squared feature selection on support vector machine text classification, in International Seminar on Application for Technology of Information and Communication (2018) 6. L. Yanfang, J. Niu, Q. Zhao, J. Lv, S. Ma, A novel text classification method for emergency event detection onsocial media, in IEEE SmartWorld, Ubiquitous intelligence computing, Advanced Trusted Computing, Scalable Computing Communications, Cloud Big Data Computing, Internet of People and Smart City Innovation (2018) 7. Deep learning, [Online]. Available: https://en.wikipedia.org/wiki/ Deep_learning 8. Artificial neural network, [Online]. Available: https:// en.wikipedia.org/wiki/Artificial_neural_network. S.J. Selvarani, Automatic identification and detection of altered, in International Conference on Intelligent Computing Applications (Coimbatore, 2014), pp. 239–243 9. Unfolded basic recurrent neural network, [Online]. Available: https://en.wikipedia.org/wiki/Recurrent_neural_network#/media/ File:Recurrent_neural_network_unfold.svg 10. J. Pennington, R. Socher, C.D. Manning, Glove: Global vectors for word representation, in Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (2014) 11. S. Minaee, N. Kalchbrenner, E. Cambria, N. Nikzad, M. Chenaghlu, J. Gao, Deep learning – Based text classification: A comprehensive review, in ACM Computing Surveys (CSUR) (2021) 12. S. Bird, E. Klein, E. Loper, Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit (O’Reilly Media, Inc, Beijing, 2009)
262 13. Word embedding, [Online]. Available: https://en.wikipedia.org/ wiki/Word_embedding 14. Naive Bayes Classifier, [Online]. Available: https:// en.wikipedia.org/wiki/Naive_Bayes_classifier 15. SVM, [Online]. Available: https://en.wikipedia.org/wiki/Supportvector_machine 16. D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning representations by back-propagating errors. Nature 323(6088) (1986) 17. P.J. Werbos, Backpropagation through time: What it does and how to do it. Proc IEEE 78(10) (1990) 18. S. Hochreiter, J. Schmidhuber, Long short-term memory. Neural Comput. 9(8) (1997) 19. LSTM, [Online]. Available: https://en.wikipedia.org/wiki/ Long_short-term_memory#/media/File:The_LSTM_Cell.svg
R. Chittari et al. 20. K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, Y. Bengio, Learning phrase representations using RNN encoder-decoder for statistical machine translation (2014) 21. Bidirectional RNNs, [Online]. Available: https://en.wikipedia.org/ wiki/Bidirectional_recurrent_neural_networks 22. R. Pascanu, C. Gulcehre, K. Cho, Y. Bengio, How to construct deep recurrent neural networks (2013) 23. D.P. Kingma, J. Ba, Adam: A method for stochastic optimization (2014) 24. N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout: A simple way to prevent neural networks from overfitting. 15(1) (2014)
Recommendation System Using MixPMF
32
Rohit Gund, James Andro-Vasko, Doina Bein, and Wolfgang Bein
Abstract
The objective of this paper is to the use of the Probabilistic Matrix Factorization (PMF) model which scales linearly with the number of observations and performs well on the large, sparse, and imbalanced music/movie dataset. In this project, we compare various PMF-based models and apply them to the recommendation system. We design and develop a Mix Probabilistic Matrix Factorization (MixPMF) model for music recommendation. This new model will take advantage of user network mapping and artist tag information and forms the effective rating matrix and thus will be efficient in recommending music/movies to new users. Simulation results show the advantage of our model. Keywords
CPMF · KPMF · Machine learning · ML · Probabilistic Matrix Factorization · PMF · Rating matrix · Music recommendation system · Music streaming service · Movie recommendation system · Video streaming service
R. Gund · D. Bein () Department of Computer Science, California State University, Fullerton, Fullerton, USA e-mail: [email protected]; [email protected] J. Andro-Vasko · W. Bein Department of Computer Science, University of Nevada, Las Vegas, Las Vegas, USA e-mail: [email protected]; [email protected]
32.1
Introduction
Recommendation System in Machine Learning (ML) context is a class of learning algorithms that offers relevant suggestions to the users. It is useful in many different applications like for music/movie recommendations. Existing approaches for recommendation systems like collaborative filtering cannot handle large datasets and cannot deal with music or movies which have very few ratings. There are over two hundred video streaming services and more than ten music streaming services available in the USA. Out of those hundreds of services, only a handful can recommend good content to its users. This makes the recommendation system extremely essential these days. A perfect recommendation system is one that can recommend the best possible options as per the users. With conventional recommendation systems like collaborative filtering, the main problem is they cannot linearly scale with data and often lack accuracy with sparse and imbalanced data. This proposed recommendation system is made using Probabilistic Matrix Factorization (PMF) [1, 2] and can deal with sparse and imbalanced data effectively. Existing available methods for recommendation systems include collaborative filtering and content-based filtering. In collaborative filtering, If a person ABC has the same opinion as a person XYZ on a topic, ABC is more likely to have XYZ’s opinion on a different topic than that of a randomly chosen person.[3]. Contentdriven filtering approaches are based on an object definition and a consumer expectations profile. These methods are best suited to situations where information about an item is known (name, location, description, etc.), but not about the user. Content-based recommendations treat recommendations as a user-specific classification problem and learn a classifier based on the functions of an item for the user’s likes and dislikes. [4] These methods do not work well when data starts
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_32
263
264
R. Gund et al.
Base Models
Dataset
DB
PMF Model
Database
DB
Generating Structured data
cPMF Model
Extracted structured data
ML Models
Database
Unique Users
Music Files
KPMF Model
Tag information
Table Storage
MixPMF Model ML Model
Recommendations Generated
Fig. 32.1 Architecture of the MixPMF recommendation system
scaling linearly. To meet this data-intensive demand PMF was introduced by Mnih in 2008 [5]. Conventional recommendation systems built using Collaborative Filtering are effective, but when it comes to imbalanced data, it gets very challenging and difficult to produce good results. To engage the users with music/movies streaming service, it is very important to recommend the content they might like and simultaneously deal with the content which has no ratings and new users onboarding on the platform. A Mix Probabilistic Matrix Factorization (MixPMF) model can deal with this problem and can generate effective music/movie recommendations. After studying the existing PMF models like PMF, constrained PMF and kernelized PMF and understanding their results, we can conclude that the additions of user network and artist tag information introduced in KPMF help us to improve the recommendation. Also, the constraint matrix introduced in CPMF helped to form a better rating matrix. We will combine these two approaches in our MixPMF model to recommend better music. The architecture of our proposed system is shown in Fig. 32.1 and has three main components: 1. Dataset “hetrec-2011” [6], 2. PMF, KPMF and CPMF models are the base models for understanding effectiveness of PMF in the recommendation system, and 3. MixPMF Model that can generate recommendations.
The paper is organized as follows. In Sect. 32.2 we present the background work, followed by our proposed model in Sect. 32.3, and model training and evaluation in Sect. 32.4. Concluding remarks and future work are given in Sect. 32.5.
32.2
Background Work
We studied three models: – Probabilistic Matrix Factorization (PMF) – Constrained Probabilistic Matrix Factorization (CPMF) – Kernelized Probabilistic Matrix Factorization (KPMF) The notations used in models are explained below: N = Number of users M = Number of items R ∈ RN × M = Rating Matrix U ∈ RN × D = User Latent Matrix V ∈ RM × D = Items Latent Matrix D = Latent Dimension W = Similarity Matrix Y = Effect of artist ratings Ku ∈ RN × M = Covariance matrix for rows of R Kv ∈ RM × M = Covariance matrix for columns of R Su ∈ RM × N = Inverse of Ku Sv ∈ RM × M = Inverse of Kv
32 Recommendation System Using MixPMF
265
Fig. 32.2 PMF diagram Fig. 32.3 CPMF diagram N×M
I∈R = Indicator matrix taking 1 if Ri, j is an observed entry, 0 otherwise. α = learning rate N(μ, σ2 ) = Normally distributed with mean μ and variance σ2 . The notation X∼N(μ, σ2 ) means that X is distributed N(μ,σ2 ) The Probabilistic Matrix Factorization (PMF) was introduced by Minh and Salakhutdinov in 2008 [5] and has shown flexibility and effectiveness in terms of large, sparse, and imbalanced data. PMF is a generative algorithm (see Fig. 32.2), and it assumes the following generative process for Rating Matrix R as follows: 1. Generate Ui ∼ N 0, σU2 I for i ∈ {1, 2, 3, · · · , N} 2. Generate Vi ∼ N 0, σV2 I for i ∈ {1, 2, 3, · · · , M} 3. For each non-missing entry of Ri, j , generate Ri,j ∼ N Ui , VjT , σ 2 I
Fig. 32.4 Additional input for KPMF model
Ui = Ui + ∝ erri,j Vj −
Vj = Vj + ∝ erri,j Ui −
Our ratings are logarithms of the user’s music listening time or video viewing time. So, we will use dot products of U and V directly. The log-posterior over the latent matrices U and V is given by following formula: log p U, V | R, σ 2 , σU2 , σV2 = log log p U, V, σ 2 + log log p U, V, σU2 + log log p V | σV2 + C where C is constant and does not depend on U and V. MAP estimate will be performed to learn U and V which maximizes by stochastic gradient descent. For each nonmissing, update rule is as follow: erri,j = Ri,j − Ui VjT
σ2 N 2
σU
p=1 Ii,p
σ2 N 2
σV
p=1 Ip,i
Ui
Vj
Mnih [5] enhanced the performance by trying to find other users who listened to the same artists as user ‘i’ and constrain the latent vector of similar users to the model and named it as CPMF (see Fig. 32.3), which contains U with latent similarity constraint matrix W ∈ RM × D [7]. KPMF models were introduced by Zhou in 2012 [8], these models allow U and V to capture the covariances between any two rows of U and V by assuming the columns of U and V are generated from zero mean Gaussian Process (GP). In KPMF models, we will use two additional inputs: user network and artist tag information (see Fig. 32.4). The generative process for KPMF model (see Fig. 32.5) is as follow:
266
R. Gund et al. Table 32.1 Stats of dataset # 1 2 3 4 5
Item User Artists Ratings Rating Density tags
Stats 1800 1000 55000 3.02% 85000
Sci-kit is used to implement the machine learning part. The proposed MixPMF model is shown in Fig. 32.6. We used the following files:
Fig. 32.5 KPMF diagram
1. Generate Ud GP(0, Ku ) for d ∈ {1, 2, 3, · · · , D} 2. Generate Vd GP(0, Kv ) for d ∈ {1, 2, 3, · · · , D} 3. For each non-missing entry of Ri, j , generate Ri,j ∼ T 2 N Ui , Vj , σ
32.3
Our Proposed MixPMF Model
We used CPMF and KPMF models for building our own variation of the PMF model and calling it the MixPMF model. The KPMF model introduced the user network information which proved to be very effective. We used the constrained vector reference from the CPMF model and combining methods of CPMF and KPMF. The steps of our model are as follows: Generate Wk ∼ N 0, σW2 I for k ∈ {1, 2, 3, · · · , M} Generate Yi ∼ N 0, σV2 I for i ∈ {1, 2, 3, · · · , N} Generate Vd GP(0, Kv ) for d ∈ {1, 2, 3, · · · , D} [9] Generate indicator matrix I such that Ii, j = 1 if Ri, j is observed or Ii, j = 0 otherwise 5. For each non-missing entry of Ri, j , generate
1. 2. 3. 4.
Ri,j ∼ N
Yi +
M k=1 Ii,k Wk M k=1 Ii,k
VjT , σ 2
Variable Description: Vd = Latent Dimension Vector W = Similarity Matrix Y = Effect of artist ratings Matrix For implementing our MixPMF model, we used NumPy and Pandas to perform the mathematical computation for data and
– MixPMF.ipynb to compute the respective matrices – Algobase function from Algobase.ipynb inherited in the cKPMF function in MixPMF.ipynb file – _compute_metrices(self, x) will calculate the rmse values after passing the Numpy array – predict_pair() will return the model rating prediction for a given user and item pair; the input to these function is the structured music dataset. These functions are then externally accessed from experiment files and predicted results are stored into text file for easy reading and understanding.
32.4
Model Training and Evaluation
We used the hetrec-2011 [6] dataset which is sparse and imbalanced to prove the effectiveness of the MixPMF model. We executed PrepData.ipynb file to generate the data files suitable for this project on the .dat file from hetrec-2011 [6]. These data files give clear information about users, items, artists, ratings and rating density, etc.. We observed a total 55000 ratings with 1800 unique users, 1000 unique artists and rating density is 3.02% while the total tags remain at 85000. Stat table, generated using the PrepData.pynb file, and shown in Table 32.1, note that the distribution of 55000 ratings are highly distorted and only 1300 ratings are larger than 5000. Due to this we opted for logarithmic count and Fig. 32.7 shows that rating distribution before and after log transformation. We note that in Fig. 32.7, only a few artists in the dataset are rated and almost 80% artists never received any ratings from users. We note that only handful users rate to music (see Fig. 32.8). The most used method for model evaluation is RMSE (root-mean-square error). This is defined as the square root of the average squared distance between the actual score and the predicted score:
32 Recommendation System Using MixPMF
267
Fig. 32.6 The proposed MixPMF model picture representation
Fig. 32.7 Frequency Distribution for Artists Fig. 32.8 Frequency distribution for users
n 1 2 yi − yˆ i RMSE = n i=1 where yi denotes the true score for the ith data point and yˆ i denotes the predicted value. This is the Euclidean distance between the vector of the true scores and the vector of the predicted scores, averaged by n, where n is the number of data points [10]. The models will inherit the PMF properties from base.ipynb file and the function AlgoBase passed as a parameter to KPMF, CPMF and MixPMF models. At the end of each iteration, the model will generate the predicted and true values using the sci-kit learn libraries inbuilt method [11]. We use an array of these values and passing
Table 32.2 RMSE value for different models Model PMF CPMF KPMF MixPMF
On 20% dataset 2.73 1.66 2.01 0.88
On 50% dataset 2.57 1.36 1.98 0.83
On 100% dataset 2.8 1.32 1.99 0.87
it as a Numpy array to calculate RMSE values. We refer to these values in model evaluation. The final calculated RMSE is shown in Table 32.2. As per the model evaluation method RMSE, we can conclude that our MixPMF model is efficient compared to PMF, CPMF and KPMF models. MixPMF model is working
268
R. Gund et al.
with RMSE of 0.83 and 0.87 on 50% and 100% of the dataset respectively.
32.5
Conclusion and Future Work
We found that combining the properties of CPMF and KPMF recommendation generations is efficient and thus we used these methods to proposed our MixPMF model. The MixPMF model is evaluated and tested with well-defined test cases. RMSE of all the models compared on the same music dataset and MixPMF model stand most reliable and efficient for both partial data from dataset and entire data from dataset. Machine learning model MixPMF is capable of recommending music efficiently than PMF, CPMF and KPMF models. As future work, developing a GUI with an automatic playlist is a future enhancement. In our project, we focused only on the back-end development. Songs generated using the MixPMF model can be analyzed for a few days and based on results they will be stored as an automatic playlist for a particular user. We also propose to investigate using a parallel processing and multithreading approach which can scale up these models to industry level applications and handles millions of files per second. Acknowledgment This research was sponsored by the NATO Science for Peace and Security Programme under grant SPS MYP G5700.
References 1. Hui Fang, Yang Bao, Jie Zhang . Leveraging Decomposed Trust in Probabilistic Matrix Factorization for Effective Recommendation, https://people.eecs.berkeley.edu/tinghuiz/papers/sdm12_kpmf.pdf 2. PMF for Recommender Systems: https://towardsdatascience.com/ pmf-for-recommender-systems-cbaf20f102f0 3. Collaborative Filtering Recommender Systems: https:// www.nowpublishers.com/article/Details/HCI-009 4. Recommendation System Wiki: https://en.wikipedia.org/wiki/ Recommender_system 5. Probabilistic Matrix Factorization – https://papers.nips.cc/paper/ 2007/hash/d7322ed717dedf1eb4e6e52a37ea7bcd-Abstract.html. Accessed on 20 May 2021 6. Music Dataset https://grouplens.org/datasets/hetrec-2011/ 7. Sources and temporal variations of constrained PMF factors obtained from multiple-year receptor modeling of ambient PM2.5 data from five speciation sites in Ontario, Canada: https://www.sciencedirect.com/science/article/abs/pii/ S1352231015001806 8. Kernelized Probabilistic Matrix Factorization: Exploiting Graphs and Side Information https://people.eecs.berkeley.edu/tinghuiz/ papers/sdm12_kpmf.pdf. Accessed on 21 July 2020 9. A Visual Exploration of Gaussian Processes: https://distill.pub/ 2019/visual-exploration-gaussian-processes/ 10. RMSE: https://vijay-choubey.medium.com/how-to-evaluate-theperformance-of-a-machine-learning-model-d12ce920c365 11. Sci-Kit RMSE: https://scikit-learn.org/stable/modules/generated/ sklearn.metrics.mean_squared_error.html
Abstractive Text Summarization Using Machine Learning
33
Aditya Dingare, Doina Bein, Wolfgang Bein, and Abhishek Verma
Abstract
Text summarization creates a brief and succinct summary of the original text. The summarized text highlights the main text’s most interesting points without omitting crucial details. There is a plethora of applications on the market that include news summaries, such as Inshort and Blinklist which not only save time but also effort. The method of manually summarizing a text can be timeconsuming. Fortunately, using algorithms, the mechanism can be automated. We apply three text summarization algorithms on the Amazon Product Review dataset from Kaggle: extractive text summarization using NLTK, extractive text summarization using TextRank, and abstractive text summarization using Seq-to-Seq. We present the advantages and disadvantages for these three methods. Keywords
Machine learning · ML · Kaggle · extractive text summarization · abstractive text summarization · text summarization · TextRank · Seq-to-Seq · NLTK · document summary A. Dingare · D. Bein () Department of Computer Science, California State University, Fullerton, Fullerton, CA, USA e-mail: [email protected]; [email protected] W. Bein Department of Computer Science, University of Nevada, Las Vegas, Las Vegas, USA e-mail: [email protected] A. Verma Department of Computer Science, California State University, Northridge, Northridge, CA, USA e-mail: [email protected]
33.1
Introduction
There are various forms of summaries: single document, multi document, informative summary, and query focused summary [1–3, 5, 7, 8]. The type of input provided to an algorithm determines these types, so for a multi-document summary, multiple documents are used. The input for query based is focused on a specific query outcome. There are two output-based primary methods for summarizing the text: abstractive text summarization and extractive text summarization [4–6, 10, 12, 13]. In extractive text summarization, the summarized text is part of the original text as the algorithm extracts the most relevant words and sentences from the original text. For example, in Fig. 33.1 the output text is consisting of all the words from the original input text only. Abstractive text summarization is opposite to extractive text summarization [7] as it returns the summary of the text that may consists of new word and sentence that are not part of the original text. For example, in Fig. 33.2, the output of the abstractive summarization consists of the words that are not part of the original text. Hybrid text summarization uses both abstractive and extractive text summarization techniques together [8]. While abstractive text summarization produces more substantive summarized text than extractive text summarization, it is more difficult to implement. Most of the research focuses on extractive text summarization’s implementation and limitations [9–11, 14, 15, 18]. We apply three text summarization algorithms on the Amazon Product Review dataset from Kaggle [12]: extractive text summarization using NLTK, extractive text summarization using TextRank, and abstractive text summarization using Seq-to-Seq. We compare their performances over various product reviews.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_33
269
270
A. Dingare et al.
Fig. 33.1 Extractive text summarization
Fig. 33.2 Abstractive text summarization
The paper is organized as follows. In Sect. 33.2 we present the background and related work, followed by the description of the algorithms used in Sect. 33.3. Simulations are results are presented in Sect. 33.4. Concluding remarks and future work are given in Sect. 33.5.
33.2
Background and Related Work
Term frequency, latent frequency, and graphical extractive algorithms are the three primary types of extractive algorithms [13, 20]. The sentence that has a similar appearance to the document word has a high score in terms of frequency. The sentences are sorted first in latent variable, and the sentence with the closest representation of the latent variable is chosen [14]. A similarity matrix is constructed in the graphical process, and the TextRank algorithm is executed based on it. The extractive text summarization can be divided into three main categories (see Fig. 33.3): term frequency, latent variable, and graphical. Term frequency and sum basic algorithm are similar: commonly occurring sentences are added together [15]. Latent variable algorithm works like as discussed above. In embedding page rank algorithm, the
embedding page vector is calculated and used during the algorithm execution [16]. Abstractive text summarization can be divided into two categories (see Fig. 33.4): semantics based, and structure based. The implementation discussed in this project is from the semantics graph-based technique. We use Text Rank algorithm developed by Mihalcea [17]. The TextRank algorithm is like PageRank PageRank algorithm that was designed and developed by Google, but instead of web pages, it uses sentences. The similarity between two sentences is the likelihood of a web page switch. This score of similarity is stored in a square matrix [18]. The standard steps in the TextRank algorithm are to load input data and construct vectors for sentences using GloVe word embeddings. The next stage is text preprocessing, which involves cleaning the data and removing common terms such as am, an, the, for, and so on. We then construct a vector representation of sentences and a similarity matrix and next we apply TextRank algorithm. The abstractive text summarization technique using Sequence-to-Sequence modeling (Seq2Seq model) is used to summarize the text. The standard implementation involves usage of encoder and decoder as referenced. The encoder and decoder are configured into two phases training and inference phase. The encoder reads input data and extracts the
33 Abstractive Text Summarization Using Machine Learning
271
Fig. 33.3 Algorithms to implement extractive text summarization
Fig. 33.4 Algorithms to implement abstractive text summarization
contextual information present in the input sequence using the Long Short-Term Memory model (LSTM). The decoder, on the other hand, uses the encoder’s output as an input and is equipped to predict the next word in the series [19]. The input data used by TextRank algorithm is a single document and does not support the use of RNNs and LSTM. It is difficult for encoder to memorize the huge data size as the fixed length vector is used to store the input data. In addition, the encoder uses a unidirectional LSTM. The context cannot be captured in both directions using a unidirectional LSTM. The bidirectional LSTM, combined with global attention for the previous problem, can be used to address the LSTM issue [20, 21].
33.3
Implemented Text Summarization Algorithms
The main steps in implementation are data gathering, data cleaning, and algorithm implementation. We implement three algorithms and compare their results.
For data gathering we used Amazon Product Review dataset from Kaggle [12] that has approximately 568,455 rows and 10 columns, almost 300 MB in size. Out of 10 columns, the ProductReviewText column has detailed description of the product available on the Amazon and is mainly used by the extractive text summarization for the TextRank algorithm. Data cleaning involves contraction mapping for handling the words with short forms like “ain’t”, “don’t” as “do not” etc., and changing the input data into either lower or upper case, remove parenthesis, eliminate stops words (e.g. “is”, “and”,” are”), punctuations, and special characters like @, # etc. These two steps are common to both abstractive and extractive algorithms [22]. The abstractive text summarization uses two columns mainly ProductReviewHeader and ProductReviewText. The column ProductReviewHeader is nothing but the header line of that particular review. This column is either one or two lines of short headline of the review.
272
A. Dingare et al.
33.3.1 Extractive Text Summarization Using NLTK We used the Natural Language Toolkit (NLTK) library for statistical language processing which include tokenization, calculating frequency of words, and calculating weighted frequency of words. Term frequency can be used to identify the keywords. Extraction of keywords in reviews enable customers in determining whether a product review is necessary and whether or not to continue reading it. Following the calculation of word frequency, a weighted frequency of each sentence in the input data column, ProductReviewText, is calculated. We calculated the frequency and weighted frequency of each word that is present in the review text. The weighted frequency can be calculated in other way by dividing the words frequency by frequency of the word that is mostly occurred [23]. The next step is to add weighted frequencies and sort the sum of weighted frequencies in descending order. The sentence with the maximum sum of weighted frequency is extracted as the summarized text. Based on the weighted frequency, the summary of the original text is returned.
33.3.2 Extractive Text Summarization Using TextRank The TextRank algorithm depends on PageRank algorithm. The probability among two words in the sentences is calculated. For the TextRank algorithm, the input reviews data will be subdivided into text units such as keywords, key phrases and the graph model is built. In this implementation, we used an undirected weighted graph; each node represents the sentence in the review text and the edges represents the relationship between them calculated using the formula [24]: WS V = (1 − d) +d∗ ΣVj∈In(Vi) eji/ΣVj∈Out (Vi) eJk WS(Vij) i
Each sentence is treated as a node in the text. There is an undirected right edge between the nodes corresponding to the two sentences if two sentences are identical. The following formula can be used to check sentence similarity [24]. Similarity Si, Sj = | Wk |Wk∈ Si &Wk∈ Sj | / log (|Si |) + log |Sj | where Si and Sj are two sentences of our product review where as Wk represents word in the sentence. The steps are: 1. Split the given text review T into complete sentences 2. Clean the data by deleting stop words, nouns, and verbs from the input data for each sentence
3. Build a candidate keyword graph G = (V, E), where V is a node collection of sentences, and then draw an edge between any two points if and only if these two sentences are linked. 4. Calculate the weight repetitively using the formula mentioned below. 5. The node weights are sorted in reverse order to obtain the most relevant T terms as candidate keywords. 6. The most significant T words are extracted from #5, marked in the original text, and then combined into a keyword if adjacent phrases are created. [25] The product reviews in column ProductReviewText is divided into small chunks of sentences only if it contains long sentences.
33.3.3 Abstractive Text Summarization Using Seq-to-Seq The encoder and decoder are needed for extractive text summarization using the Seq-to-Seq model. In this implementation, the Long Short-Term Memory (LSTM) is used as encoders and decoders to catch the phrase dependencies in a sentence’s words. To implement encode and decoder,” Recurrent Neural Networks i.e. RNN” can also be used. The encoders and decoders are designed further in two stages, namely training and inference. The encoder receives the input data as input and extracts the contextual information present in the data. The timestamp is also important factor here. So, for each time stamp, each word from the product review sentence is given to the encoder which retrieves the contextual information from the product review. This is the part of the training phase. Before feeding the target sequence into the decoder, their start and end are inserted [23]. The encoder training phase is single direction (see Fig. 33.5). The encoder receives the input word by word at each time interval and each steps of the LSTM passes the output to the next cell in the LSTM. The decoder receives the hidden state (hi ) and cell state (ci ) from the previous steps as data. The decoder training is single direction (see Fig. 33.6) This decoder cell also takes input from the encoder and process each word. There are total of three cells of each encoder and decoder. To calculate the past as well as future context for each product review sequence we used a bi-directional LSTM. The bidirectional LSTM in which the original cells as well as it’s transpose are used. The single-directional LSTM has the disadvantage of being unable to predict future background information and learning all the input sequence sentences. As referenced in the Fig. 33.7, in bi-directional LSTM the output of the last cell is given back to additional LSTM.
33 Abstractive Text Summarization Using Machine Learning
Fig. 33.5 Training encoder
Fig. 33.6 Training decoder [7]
Fig. 33.7 Bi-directional LSTM
273
274
A. Dingare et al.
Fig. 33.8 NLTK output
LSTM0’, LSTM1’, and LSTM2’ are all the transpose of original LSTM0, LSTM1, and LSTM2 [23].
33.4
subsequent input data. The main disadvantage of using Seqto-Seq for abstractive text summarization is that sentences are not ranked like text summarization. This will cause us to skip over text that appears frequently in the input.
Simulations and Results
For both the abstractive and extractive text summarization we considered reviews that have at least 250 words. The output of the extractive text summarization using NLTK applied to a sample product review and its German language translation of the summary (cross language summary) is shown in Fig. 33.8. The output by the TextRank algorithm of the sample product review is shown in Fig. 33.9. There are approximately more than 5 reviews being summarized and their associated German translations. Figure 33.10 shows the output of the Seq-to-Seq implementation with original text as well as summarized text. The following limitations have been noted. The extractive text summarization implemented using TextRank algorithm does not return proper output for the duplicate words and sentences. The algorithm is modified to perform the multi document text summarization. It means the algorithm checks for the .CSV files in input directory and picks up all the file while processing. However, both the TextRank and Seq-toSeq algorithms do not remember the input data from one source input file while processing another input file. In short, even though the multi document processing is supported, the inter dependency of input is not taken into consideration. Another drawback of our implementation is that the models we developed are unable to generate new product feedback that could be used in conjunction with summarizing the
33.5
Conclusion and Future Work
In this paper we apply three text summarization algorithms on the Amazon Product Review dataset from Kaggle [12]: extractive text summarization using NLTK, extractive text summarization using TextRank, and abstractive text summarization using Seq-to-Seq. There are advantages and disadvantages to using these algorithms for product reviews summarization. As future work, we note that the TextRank algorithm we used for extractive text summarization does not endorse Recurrent Neural Networks (RNN). The RNN is the most common algorithm for dealing with a continuous stream of data. Internal memory in the RNN aids in remembering the input and makes it ideal for deep learning problems. We could use RN to improve the current algorithm processing because the RNN remembers the import part of the input and uses it as a guide in subsequent runs. The RNN’s performance summarized text looks more like text summarized by a person. However, the RNN has its own disadvantage – It fails for complex model. Its output is also poor if the input text includes duplicate words and sentences. For both the abstractive and extractive text summarization, using certain user parameters, the output of the summarized text may be further refined.
33 Abstractive Text Summarization Using Machine Learning
275
Fig. 33.9 TextRank output
Fig. 33.10 Seq-to-Seq output Acknowledgment This research was sponsored by the NATO Science for Peace and Security Program under grant SPS MYP G5700.
References 1. https://www.analyticsvidhya.com/blog/2020/12/understandingtext-classification-in-nlp-with-movie-review-example-example/ 2. A Gentle Introduction to Text Summarization in Machine Learning. https://blog.floydhub.com/gentle-introduction-to-textsummarization-in-machine-learning/ 3. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, I. Polosukhin, Attention Is All You Need. [online] arXiv.org (2021). Available at: https://arxiv.org/abs/1706.03762. Accessed 17 May 2021
4. Pranab Ghosh, Six Unsupervised Extractive Text Summarization Techniques Side by Side [Online]. Available: https:// pkghosh.wordpress.com/2019/06/27/six-unsupervised-extractivetext-summarization-techniques-side-by-side/. Accessed 17 May 2021 5. J.N. Madhuri, R. Ganesh Kumar, Extractive text summarization using sentence ranking, in 2019 International Conference on Data Science and Communication (IconDSC), 2019, pp. 1–3. https:// doi.org/10.1109/IconDSC.2019.8817040 6. C. Lakshmi Devasena, M. Hemalatha, Automatic text categorization and summarization using rule reduction, in IEEE-International Conference on Advances in Engineering, Science and Management (ICAESM –2012), 2012, pp. 594–598 7. https://www.analyticsvidhya.com/blog/2019/06/comprehensiveguide-text-summarization-using-deep-learning-python/
276 8. S. Selvarani, S. Jebapriya, R. Smeeta Mary, Automatic identification and detection of altered fingerprints, in 2014 International Conference on Intelligent Computing Applications, (2014), pp. 239–243 9. R. Mishra, V.K. Panchal, P. Kumar, Extractive text summarization – An effective approach to extract information from text, in 2019 International Conference on Contemporary Computing and Informatics (IC3I), 2019, pp. 252–255. https://doi.org/10.1109/ IC3I46837.2019.9055636 10. S.M. Meena, M. P. Ramkumar, G.S.R. Emil Selvan, Text summarization using text frequency ranking sentence prediction, in 2020 4th International Conference on Computer, Communication and Signal Processing (ICCCSP), 2020, pp. 1–5, https://doi.org/ 10.1109/ICCCSP49186.2020.9315203 11. X. You, Automatic summarization and keyword extraction from web page or text file, in 2019 IEEE 2nd International Conference on Computer and Communication Engineering Technology (CCET), 2019, pp. 154–158. https://doi.org/10.1109/ CCET48361.2019.8989315 12. Consumer Reviews of Amazon Products. https://www.kaggle.com/ datafiniti/consumer-reviews-of-amazon-products 13. Y. Chen, Q. Song, News text summarization method based on BART-TextRank model, in 2021 IEEE 5th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC), 2021, pp. 2005–2010. https://doi.org/10.1109/ IAEAC50856.2021.9390683 14. https://www.analyticsvidhya.com/blog/2018/11/introduction-textsummarization-textrank-python/ 15. Abstractive Text Summarization Using Transformers | by Rohan Jagtap | The Startup | Medium. https://medium.com/ swlh/abstractive-text-summarization-using-transformers3e774cc42453 16. D. Suleiman, A. Awajan, Deep Learning Based Abstractive Text Summarization: Approaches, Datasets, Evaluation Measures, and Challenges (2021) 17. R. Mihalcea, P. Tarau, Textrank: Bringing order into texts, in Proceedings of EMNLP 2004, ed. by D. Lin, D. Wu, (Association
A. Dingare et al.
18.
19.
20. 21. 22.
23.
24.
25.
for Computational Linguistics, Barcelona, Spain., 2004), pp. 404–411 N. Patel, N. Mangaokar, Abstractive vs extractive text summarization (output based approach) – A comparative study, in 2020 IEEE International Conference for Innovation in Technology (INOCON), 2020, pp. 1–6, doi: https://doi.org/10.1109/ INOCON50539.2020.9298416 A. Rahman, F.M. Rafiq, R. Saha, R. Rafian, H. Arif, Bengali text summarization using TextRank, fuzzy C-means and aggregate scoring methods, in 2019 IEEE Region 10 Symposium (TENSYMP), 2019, pp. 331–336. https://doi.org/10.1109/ TENSYMP46218.2019.8971039 Sequence to Sequence Learning with Neural Networks arXiv:1409.3215v3 [cs.CL] 14 Dec 2014 Get To The Point: Summarization with Pointer-Generator Networks. https://arxiv.org/abs/1704.04368 S.R. Manalu, Stop words in review summarization using TextRank, in 2017 14th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON), 2017, pp. 846–849. https://doi.org/10.1109/ ECTICon.2017.8096371 W. Jiang, Y. Zou, T. Zhao, Q. Zhang, Y. Ma, A hierarchical bidirectional LSTM sequence model for extractive text summarization in electric power systems, in 2020 13th International Symposium on Computational Intelligence and Design (ISCID), 2020, pp. 290– 294. https://doi.org/10.1109/ISCID51228.2020.00071 T. Behere, A. Vaidya, A. Birhade, K. Shinde, P. Deshpande, S. Jahirabadkar, Text summarization and classification of conversation data between service chatbot and customer, in 2020 Fourth World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4), 2020, pp. 833–838. https://doi.org/10.1109/ WorldS450073.2020.9210289 M.R. Ramadhan, S.N. Endah, A.B.J. Mantau, Implementation of textrank algorithm in product review summarization, in 2020 4th International Conference on Informatics and Computational Sciences (ICICoS), 2020, pp. 1–5. https://doi.org/10.1109/ ICICoS51170.2020.9299005
Intelligent System for Detection and Identification of Ground Anomalies for Rescue
34
Antonio Dantas, Leandro Diniz, Maurício Almeida, Ella Olsson, Peter Funk, Rickard Sohlberg, and Alexandre Ramos
Abstract
The search and identification of people lost in an emergency is a very important activity, it is carried out to assist in human lives in danger, and the unavailability of support technologies. Unmanned aerial vehicles (UAV) act directly in this activity for greater capacity in the coverage of the area in a shorter time of operation. A brief state-of-the-art survey is presented, and specific needs are raised for accuracy and practical application. In this joint work, low-processing image recognition alternatives for use in UAV will be explored to face challenges such as location in large areas, small targets, and orientation. Assessments are performed on real images using Inception, SSD, and Yolo. For tracking, MIL, KCF, and Boosting techniques are applied to empirically observe the results of test missions. Preliminary results show that the proposed method is viable for action in the search and rescue of people. Keywords
UAV · Search · Rescue · Images · Recognition
A. Dantas · L. Diniz · M. Almeida · A. Ramos () Institute Mathematics Computing (IMC), Federal University of Itajubá (UNIFEI), Itajuba, Brazil e-mail: [email protected]; [email protected]; [email protected]; [email protected] E. Olsson Saab Group and IDT, Mälardalen University, Västerås, Sweden e-mail: [email protected] P. Funk · R. Sohlberg Artificial Intelligence Group, Mälardalen University, Västerås, Sweden e-mail: [email protected]; [email protected]
34.1
Introduction
Capturing images of the earth’s surface is increasingly gaining importance, particularly for the analysis of anomalies in the soil for various purposes, such as geospatial studies, fires, natural disasters, agriculture, etc. One of the most relevant scenarios is the search and identification of people who are lost or at risk, they are called search and rescue (SAR) activities, and today they are now commonly assisted by information technology [1]. Brazilian and Swedish researchers have been developing methods, techniques and tools for interpreting soil images obtained by sensors embedded in unmanned aerial vehicles (UAV), in order to help and complement the efforts of human researchers. Motivated mainly by the reduction of costs and efficiency, provided by the versatility in the process of photos captured in real time in strategic positions, in comparison with the high cost and sometimes of low efficiency, when performed in areas with few professionals or photos of satellites [2]. These systems of support for search and rescue activities, known as search and rescue unmanned aerial vehicles (SARUAV), have different obstacles to be faced in the current scenario. Adverse conditions stand out in certain regions, the area covered by the UAV camera and the change in lighting, are imposed for high-performance human detection [3]. Supporting information such as direction, speed and action can also help rescuers reach the victim more quickly and effectively [4]. In this work will present alternatives to face these challenges, for that, similar works will be presented in the recent literature, after an overview of the techniques that can be applied, then the methodology adopted in the study.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_34
277
278
34.2
A. Dantas et al.
Related Works
Presented here are some recent works related to SARUAV, presenting its strengths, as well as the improvements recommended by the authors. In [4–7] the focus of the studies is on the planning of individual flight paths and resource allocation, for which the probability of targeting is performed, proportionally, avoiding redundancy in the coverage, thus correctly allocating the available resources. These methods have a support center for optimization, and people specialized in the identification of targets. There are works in the literature using several different ways to identify people in danger. Noteworthy are the works that use artificial intelligence, in [8] implemented a system in real time, using the deep convolutional neural network of the Single Shot Detection (SSD) binary output topology, based on TensorFlow Lite, reaching a recall of 65.4 and 96.4% accuracy in images at a rate of 11 to 17 FPS, captured from a sequential search from a defined area. The HERIDAL dataset [9] was used, which has 68,750 images, 29,050 with positive examples. The author highlights the need to adjust the parameters and topology of the network, so that objects that are proportionally very small in relation to the entire captured image can be safely detected. The work of [10] with a focus on maritime rescue, also uses convolutional networks with binary output, however the input is through data from a multispectral camera, with the filter channels blue (B), red (R), green (G), red edge (REG) and near IR (NIR), which makes the cost of the project high. The results obtained by them show that the best classification performance of 89% through the Xception and SqueezeNet topology, in images captured with about 6 FPS, was obtained by combining the G, REG and NIR channels combined with the use of sliding windows. Finally in the paper [11] a lightweight system was proposed using the You Only Look Once version 3 (YOLOv3) algorithm, focusing on human detection in low-pixel (640 × 512) multispectral infrared and thermal images, that is, with use of high-cost special cameras. The results showed an accuracy of 89.26% in images at 24.6 FPS. Recent work on this topic shows that despite good results, the cost of processing, such as material and maintenance, is high. In addition, the operating scenario and path and search mode can be optimized.
34.3
Methodology
This section will present general techniques commonly used for problems in identifying anomalies in images for the most diverse purposes. The raw image, as it comes from the UAV, can be difficult to identify anomalies in the soil for both
humans and algorithms, to circumvent this image treatment techniques are employed. In this pre-processing, segmentation, resampling and cutting of the region of interest are performed, equalization and enhancement filters can also be applied, these processes and image manipulation can result in less computationally expensive processing in the classifiers [12]. Machine learning techniques are commonly used in image processing to identify subtle differences [13]. Convolutional neural networks (CNN) are a variation of the multi-layer perceptron networks (MLP) of the feed-forward type, which in turn are inspired by the biological processes of the pattern of connectivity between neurons, originating in the organization of the animals’ visual cortex. It stands out because it has been successfully applied in the processing and analysis of digital images, requiring a minimum level of pre-processing when compared to other traditional image classification algorithms, since its structure allows it to correctly recognize and apply the standardized characteristics filters taken from the training images [14]. Although it is already widely used in this subject, it has recent technologies of low operational cost for use in UAV that can be applied in this context. In the Inception V2 architecture two 3 × 3 convolutions are used, the module’s resource banks are expanded instead of deeper. This would prevent loss of image information and decrease computational time in convolution. The Single Shot MultiBox Detector (SSD) is a popular algorithm, which aims in a single shot to detect multiple objects within the image, keeping the concept of regions, unlike other approaches that need two shots, so it maintains its accuracy with less computational time. You Only Look Once (YOLO) introduced a faster way of real-time object detection, instead of using the region selection method, YOLO uses the Convolutional Neural Network which provides the bounding boxes as well as the class to these boxes, dividing the image into a constant value grid. When classified, a tracking can be performed to monitor the detected anomaly, this ensures that a certain item is closer to something alive, as it moves. In addition, when obtaining the plotted path, it is possible to make estimates of the location at a given time. With the confirmation of the target, the processing is also lighter as it does not need to classify continuously. Boosting is based on AdaBoost that acts in haar cascade, has stability and good detection despite the adjustment time. The Multi-Instance Learning (MIL) tracker is one of the most used algorithms, its principle is to look in a small neighborhood around the current location to generate several potential positive examples, where only one image can already be considered, it has a performance and even under partial occlusion. The Kernelized Correlation Filters (KCF)
34 Intelligent System for Detection and Identification of Ground Anomalies for Rescue
algorithm, its principle is similar to MIL, however it applies several mathematical properties in the overlapping regions of positive samples for optimization, and therefore obtains better precision and processing speed [15]. To develop a decision support system in the search and rescue of people, it is necessary to approach the reality of the professionals involved. Thus, it was proposed to create a dataset with the most different situations. The images were obtained from a DJI Mavic drone with a simple 1/2.3 camera (CMOS), up to 12.35M pixels, and 78.8◦ 26 mm field of view (FOV) lenses, Distortion , emanating from a point in Chi , there is a corresponding visibility ray r2 = [x,b>, emanating from a point x in the lid and coinciding with r1 . This implies that the part of VP2 outside the lid is identical to VP1 . C-Chain Approximation The approximation algorithm we present is based on identifying concave chains of ε-width (referred now on simply as ε-C-Chains) on the boundary. We can now describe the C-Chain Approximation algorithm. The inputs to the algorithm are (i) error tolerance level ε, and (ii) a 1.5D terrain T. The algorithm outputs approximated terrain T’ whose vertices are within ε distance from some point on the boundary of T. The algorithm scans the boundary of T and identifies ε-C-Chains one at a time such that the identified ε-C-Chains are disjoint from each other. For capturing ε-C-Chains only those chains are considered that form concave a chain from outside. The boundary of 1.5D Terrain is partitioned into maximal concave chains as shown in Fig. 45.5, where the lids of the maximal concave chains are drawn red. For the purpose of describing the algorithm we take the length of the edge (Fig. 45.5a) as the value of the error tolerance level ε. The widths of some maximal concave chains are larger than ε. For example, the width of maximal concave chain is far larger than ε. Similarly, the width of the maximal concave chain is bigger than ε. Large width maximal concave chains are partitioned into the minimum number of concave chains whose widths are less or equal to ε we obtain the intended sequence of concave chains. The approximation of the terrain’s boundary is achieved by replacing each concave chain with their lids which is shown in Fig. 45.5c. In Fig. 45.5c, the approximated boundary is drawn with dashed edges, where the original boundary with 35 vertices is approximated by a boundary with 24 vertices. We now proceed to develop a d formal sketch of the C-Chain Approximation algorithm. The proposed C-Chain Approximation algorithm makes use of the function for extracting ε-C-Chains from a given concave chain CH = . The function examines the width of sub
383
chains CHi,j = of lengths 3, 4, 5, ... one at a time to find the consecutive sub chains CHi,j and CHi,j+1 such that the conditions listed below are satisfied. In these conditions, WD(CHi,j ) denotes the width of the chain CHi,j = . WD CH i,j < ε
(45.1)
WD CH i,j+1 > ε
(45.2)
We refer to the first vertex of the sub-chain vi as first leg and the end vertex vi as last leg. The algorithm initially starts with sub-chain CHi,i+2 (the triangle corresponding to the first three vertices) as the candidate ε-C-Chain. If the width of CHi,j = CHi,i+2 is greater than ε then the input chain has no ε-CChain and hence CHi,j cannot be approximated. On the other hand, if the width of CHi,i+2 is smaller than ε then the second leg (j) is incremented and the width of CHi,j is checked. This process of moving the second leg continues until the end of the input chain is reached or conditions (3.1) and (3.2) are satisfied. If the end of the input chain is reached, then the last ε-C-Chain is found. If conditions (3.1) and (3.2) are satisfied without reaching the end of the input chain, then a ε-C-Chain is found and the algorithm proceeds to identify the next ε-CChain by resetting the first leg to the running last leg and last leg equal to new first leg plus 2. A formal description of the algorithm for extracting ε-CChain components for a given input maximal concave chain is listed as Algorithm 45.1. Algorithm 45.1: Extract ε-C-Chains 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15:
}
Input: (i) Concave chain ch1 = Input: (ii) Tolerance threshold ε Output: Disjoint ε-C-Chains components i=1; j = 3; while (j≤n) do { if ( wd(ch1 , i, j) ≥ ε ) i++; j = j + 2; else if (j≥n ) Output segment (pi , pj ); else if (wd(ch1 , i, j) ≥ ε) { Output segment (pi , pj ); i = j; j = j +2; } elsej++;
The actual C-Chain approximation algorithm works by first marking maximal concave chains on the boundary of 1.5D terrain T. The consecutive sequence of vertices with external angle greater than 180 degrees are precisely the maximum concave chains. All such chains can be marked
384
L. Gewali and S. Jha
Fig. 45.5 Alternate sequence of concave/convex chains
in one scan of the boundary by checking external angles. In the second pass around the boundary of T, Algorithm 1 is invoked for each maximal chain. Replacing each ε-C-Chain with corresponding lids gives the approximated boundary T’. Time Complexity The time complexity of the proposed algorithm can be analyzed in straightforward manner. Algorithm 1 checks each vertex constant number of times. The shape formed by the maximal concave chain and its lid is a convex polygon such that the projection of edges of the concave chain on the lid are disjoint. Because of convexity and the disjoint property, the width of the shape corresponding each ε-C-Chain can be computed in time proportional to its size. Hence the total time to build the approximated chain is O(n) where n is the number of vertices on the boundary of 1.5D terrain T.
45.4
Discussions
The problem of boundary approximation subject to external visibility has rich applications in (i) the placement of cellular tower and (ii) covert path planning for surveillance on terrain
surface. If the input 1.5D terrain is of very large size, then the proposed approximation algorithm can be applied to obtain a reduced size 1.5D terrain. Furthermore, in situations where we need to transmit the terrain data we could apply the proposed approximation algorithm to compress the input and transmit efficiently. For applications in covert path planning the objective is to construct a collision-free path over the 1.5D terrain such that the constructed path is least exposed to observation tower placed on the surface. A covert path P should be such that its exposure to points on terrain surface should be reduced and at the same time the path should have acceptable clearance from the terrain. Somehow we need to extrapolate visibility exposure of regions above the terrain surface to construct such a path. A path that penetrates the regions of low exposure can be taken as a covert path. One approach would be to evaluate external visibility profile of region outside the terrain surface by extending the visibility rays induced by the edges of 1.5D terrain. If we extend the rays emanating from the edges of the terrain we get an arrangement of visibility rays as shown in Fig. 45.6, where visibility rays are drawn with dashed lines. A path above the ground that crosses the minimum number of visibility edges could be taken as a covert path. A path
45 Boundary Approximation and External Visibility
385
Fig. 45.6 Illustrating a possible covert penetration
consisting of line segments as shown by red colored edges in Fig. 45.6 could be taken as the initial candidate for covert path. Some of the turn angles implied by the initial path may be too sharp for an aerial vehicle used for reconnaissance mission. To address such problems sections of the path (subpaths) near the sharp turns need to patched. For example, in Fig. 45.6, the part of the path above the terrain vertices v15 and v16 has a very sharp turn which is clearly forbidden for aerial vehicles. Such portions of the paths could be patched by a detour path that does not have sharp turn and its visibility to other parts of the terrain is reduced. Details of constructing covert path above a terrain are in progress and will be reported in the future. For computing ε-C-Chains (in Sect. 45.3) we required that the external visibility from the C-Chain is identical to external visibility from the corresponding approximating segment. We can relax this requirement to develop another version of approximation algorithm described in Sect. 45.3. The detail of this version of the algorithm would be worth investigating.
References 1. W.S. Chan, F. Chin, Approximation of polygonal curves with minimum number of line segments or minimum error. Int. J. Comput. Geom. Appl. 6, 59–77 (1996) 2. M. de Burg, M. van Kreveld, M. Overmars, O. Schwarzkopf, Computational Geometry: Algorithms and Applications, 2nd edn. (Springer, 2000) 3. D. Douglas, T. Peucker, Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica Int. J. Geogr. Inf. Geovisualiz. 10, 112–122 (1973)
4. S. Eidenbenz, Approximation algorithms for terrain guarding. Inf. Process. Lett. 82(2), 99–105 (2002) 5. L. Gewali, S. Jha, Effect of boundary approximation on visibility, (in ITNG 2021). Adv. Intell. Syst. Comput. 1346, 247–253 (2021) 6. S.K. Ghosh, D.M. Mount, An output sensitive algorithm for computing visibility graphs, in 28th Annual Symposium on Foundations of Computer Science (sfcs 1987), pp. 11–19 7. M. Hacar, T. Gokgoz, A new score-based multi stage matching approach for road network conflation in different road patterns. ISPRS Int. J. Geo Inf. 8(2), 81 (2019) 8. H. Imai, M. Iri, Computational-geometric methods for polygonal approximations of a curve. Comp. Vis. Graph. Image Process 36(1), 31–41 (1986) 9. J. O’Rourke, Computational Geometry in C, 2nd edn. (Cambridge University Press, Cambridge, 1998) 10. Z. Xie, Z. Ye, L. Wu, Research on building polygon map generalization algorithm, in Eighth ACIS International Conference on Software Engineering, Artificial Intelligence, Networking, and Parallel/Distributed Computing, vol. 3 (SNPD, 2007), pp. 786–791
Detection of Strictly L3-Live Structures by Structural Analysis of General Petri Net Using SAT-Solver
46
Yuta Yoshizawa and Katsumi Wasaki
Abstract
One of the dynamic properties of Petri nets is liveness, which ranges from L0 to L4 depending on the severity of the condition. The structure required for the existence of strictly L3-live transitions is referred to as the strictly L3-live structure. The strictly L3-live structure consists of three elements: a repeating closed circuit (L3-circuit), a transition that cancels the liveness of the L3-circuit (CircuitBreaker), and a place that receives a token supplied by the L3-circuit (k-place). To detect these elements, we use the partially-conservative, partially-repetitive, and bounded structurally properties of general Petri nets, as well as solving the matrix inequalities, which are necessary and sufficient conditions for the three structural properties, using the SAT solver to detect the three elements that constitute the strictly L3-live structure. Keywords
Petri nets · Model check · Liveness check · SAT solver · HiPS tool · Structural properties · Strictly L3-live · Strictly L2-live · Structural analysis · Liveness level
46.1
Introduction
Recently, information technology-based systems have been used in various aspects of our lives, and the scale and numY. Yoshizawa Graduate School of Science and Technology, Shinshu University, Nagano, Japan e-mail: [email protected] K. Wasaki () Faculty of Engineering, Shinshu University, Nagano, Japan e-mail: [email protected]
ber of these systems have been increasing. Because these technologies serve as the foundations for social systems, ensuring not only the performance but also the reliability of these systems has become critical. However, guaranteeing the design and operation of parallel systems, such as asynchronous circuits and asynchronous communication protocols, is difficult because errors do not easily appear during verification. One way to verify the reliability of these parallel systems is to model them with Petri nets. Petri nets are mathematical models that show the behavior of discrete event systems with parallelism, asynchrony, and non-determinism of event occurrence [1, 2]. Petri nets can perform three functions simultaneously: graphical tools, simulation tools, and mathematical methodologies. This describes the system structure in a visual representation as a graphic tool, and it can simulate parallel and dynamic events in the system using tokens in Petri nets. However, as a mathematical tool, it can build equations of state, algebraic equations, and other models that describe the behavior of the system. A Petri net support tool, HiPS (Hierarchical Petri Net Simulator), has been developed at Shinshu University to describe and analyze Petri nets [3–5]. HiPS has an intuitive and common GUI and various functions such as structural and dynamic property analysis. Petri nets represent a system’s simulation by manipulating the firing of transitions. The possibility of a transition firing is referred to as activity level, and it is closely related to the system being deadlocked or not happening at all. Being active is an ideal property of many systems, but it is a demanding property for large systems; thus, activity levels have been defined, with the activity condition relaxed at each level. Understanding the behavior of a system requires knowing the activity level a transition in a Petri net belongs to. In this study, we describe the structure and analysis required for the existence of activity level-3 transitions. Here, we discuss the structural properties of general Petri nets and how to SAT-solver to efficiently analyze them. Section 46.2
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_46
387
388
Y. Yoshizawa and K. Wasaki
describes the concept and properties of Petri nets. Section 46.3 defines the strictly L3-live structure. Section 46.4 describes the structure detection method. Section 46.5 presents examples of structural analysis.
46.2
j, and A+ =
a+ ij is called the positive incidence matrix.
Furthermore, a− the ij = w (j, i) is the weight of the arc from − input place j of transition i to transition i, and A− = aij is known as the backward connection matrix.
Petri Nets 46.2.3 Firing Condition
46.2.1 Place/Transition Net Carl Adam Petri proposed the Petri net model to graphically represent a discrete event system consisting of multiple processes. It can be used to describe and study information processing systems with asynchronous, distributed, parallel, nondeterministic, and stochastic behavior. The model can be used to describe and study information processing systems with linear, asynchronous, distributed, parallel, nondeterministic, and stochastic behavior. It can be used to simulate system behavior using tokens in Petri nets, additionally to be used as a visual representation of the system structure. A Petri net is a directed bipartite graph whose vertices are places and transitions, and whose four components are arcs and tokens. In a Petri net, places are drawn as circles, and transitions are drawn as squares or bars. Arcs connect places to transitions and transitions to places. They are represented by arrows according to their connection directions. A P/T-net is a 5-tuple, PN = (P, T, F, W, M0 ), where: P = {p1 , p2 , · · · , pm } is a finite set of places, T = {t1 , t2 , · · · , tn } is a finite set of transitions, F ⊆ (P × T) ∪ (T × P) is a set of arcs (flow relations), W : F → {1, 2, 3, · · · } is a weight function, M0 : P → {0, 1, 2, 3, · · · } is the initial marking, P ∩ T = ∅, and P ∪ T = ∅. A Petri net structure N = (P, T, F, W) without any specific initial marking is denoted by N. A Petri net structure N = (P, T, F, W) without any specific initial marking is denoted by N. For a Petri net (N, M0 ), the set of all possible markings reachable from the initial marking M0 is denoted by R(N, M0 ) or simply R(M0 ). The set of all possible firing sequences from M0 in a net (N, M0 ) is denoted by L(N, M0 ) or simply L(M0 ).
46.2.2 Incidence Matrix In a Petri net (N, M0 ) such that the number of transitions is n and the number of places is m, the The incidence matrix A = [aij ] is an n -by- m matrix that represents the connection between transitions and places. Each component of the inci− + dence matrix is given byaij = a+ ij − aij . Here, Aij = w (i, j) is the weight of the arc from transition i to its output place
The transition t can fire if there are more tokens in the input place p of the transition t than the weight of the arc from the input place p to the transition t. The transition t is firable if the input place p has at least one token of the weight of the arc from the input place p to the transition t. A firable transition can either fire or not fire. This is because, while an event may be in a state where it can occur, it may or may not occur. When the transition t fires, a token of the arc weight is removed from each input place. When the transition t fires, a token of the arc weight is removed from each input place. Then, a token for the weight of the arc is added to each output place of t.
46.2.4 Structural Properties Structural properties are Petri net properties that are either independent of the initial marking or associated with the presence of a particular firing sequence from some initial marking. These properties can be described by the incidence matrix A and its associated homogeneous equations and inequalities because they depend on the structure of the Petri net. In the necessary and sufficient condition equation, x denotes the firing frequency vector, and y denotes the weighted sum of the total number of tokens. 1. Structurally Bounded (SB): A Petri net N is said to be structurally bounded if it is bounded for any finite initial marking M0 . The necessary and sufficient conditions are
∃y > 0, Ay ≤ 0
(46.1)
2. Partially Repetitive (PRP): A Petri net N is said to be partially repetitive if there exists a marking M0 and a firing sequence σ from M0 such that some transitions occur infinitely. The necessary and sufficient conditions are
∃x ≥ 0 (x = 0) , AT x ≥ 0
(46.2)
3. Partially Conservative(PCS): A Petri net N is said to be partially conservative if there exists a positive integer
46 Detection of Strictly L3-Live Structures by Structural Analysis of General Petri Net Using SAT-Solver Table 46.1 Liveness level L1-live The transition t can fire at least once in the firing sequence with L(M0 ). L2-live For any positive integer k, the transition t is firable at least k times in some firing sequence of L(M0 ). L3-live The transition t appears infinitely many times in the firing sequence with L(M0 ). L4-live Transition t is L1-live for all markings in R(M0 ).
solution y(p) for some place p, such that the weighted total number of tokens (M T y = M0T y) is constant for all M ∈ R(M0 ) and any fixed initial marking M0 . The necessary and sufficient conditions for partially conservative are
∃y ≥ 0 (y = 0) , Ay = 0
(46.3)
46.2.5 Liveness One of the dynamic properties of Petri nets is a property known as Liveness. We say a Petri net (N, M0 ) is Liveness if any transition in the net can be fired from any marking M0 through any firing sequence, regardless of which marking is reached from the initial marking M0 . This means that an active Petri net ensures deadlock-free operation regardless of the firing sequence used. Therefore, while liveness is an ideal property for many systems, verifying this stringent property is prohibitively expensive for large systems. Therefore, the levels of liveness shown in Table 46.1 are defined by relaxing the liveness condition [7, 8]. A Petri net (N, M0 ) is said to be Lk-live if all transitions in the net are Lk-live (k = 1, 2, 3, 4). The L4-live is the strongest activity level and corresponds to the liveness defined. A transition is said to be strictly Lk-live if it is Lk-live and not L(k + 1)-live, where (k = 1, 2, 3). In the following, when we say Lk-live or Lk-live, we mean strictly Lk-live.
46.3
389
can fire, t1 has already fired and t3 is no longer active, so the upper limit of the number of tokens in p2 is bounded. Therefore, while t2 cannot fire an infinite number of times in any firing sequence, it can fire as many times as t3 fires, making it strictly L2-live. Therefore, the existence of strictly L2-live transitions depends on the existence of strictly L3-live transitions. In order for strictly L3-live transitions to exist, there must be a repeating closed circuit (L3-circuit) such as the one composed of t3 and p1 , and a transition (CircuitBreaker) such as t1 that removes the liveness of the circuit. The strictly L3-live structure is composed of these three elements plus a place (k-place) that receives a token from a L3circuit transition such as p3 . The existence of L3-circuit and CircuitBreaker is required for the strictly L3-live structure to exist. (a) L3-circuit: A Petri net circuit has a structure in which all transitions on the circuit are iterative and deprive them of activity. In the net in Fig. 46.1, the circuit consisting of t3 and p1 is repetitive and its liveness is taken away by the firing of t1 , resulting in a L3-circuit. (b) CircuitBreaker: A transition that removes the liveness of a transition in a L3-circuit that has a place in the L3circuit as input; the input place in the L3-circuit must be a conservative place, and the transition must be strictly L1-live. This transition is represented by t1 in the net in Fig. 46.1. (c) k-place: A place outside the L3-circuit receives tokens from transitions inside the L3-circuit and is infinitely supplied with tokens only by transitions inside the L3circuit. An extended k-place is a place that receives tokens from a k-place and has the same properties as a k-place. p2 is a k-place in the net in Fig. 46.1.
46.4
Structural Analysis
The flow of the structural analysis is shown in Fig. 46.2, where the structural properties are determined using the SATsolver, and each structure is analyzed by net exploration using the solution. The analysis method is explained in Fig. 46.3.
Proposal of Srictly L3-Live Structure
The strictly L3-live structure is the structure required for the existence of strictly L3-live transition. t3 is strictly L3-live in the net of Fig. 46.1 because it can fire an infinite number of times if t1 does not fire. After t1 has fired, t2 can fire as many times as t3 has fired (the number of tokens in p2 ). When t2
46.4.1 Solving structural properties using SAT-solver SAT-solver is used to obtain the necessary and sufficient conditions for structural properties. Google OR-Tools [6] is used as the solver. When solving for structural partially
390
Y. Yoshizawa and K. Wasaki
Fig. 46.1 Contain L3 transition net
Fig. 46.2 Flow of structural analysis
repetitive, the elements of the array of solutions obtained are converted to bool values, and a conditional expression is added to exclude solutions with the same bool values as those already obtained when solving for the next solution.
{t2 , t3 }. Additionally, t4 , a non-repetitive transition with a conservative p2 as input, is determined to be a CircuitBreaker, and the set of conservative places {p0 , p1 , p2 }, as well as its input place p2 is saved as CircuitBreakerSource.
46.4.2 Detection of CircuitBreaker
46.4.3 Detection of L3-circuit
We can determine which places are conservative by finding a solution to structural partially conservative. Determine that the non-iterative transition whose input is a place on the iterative circuit is a CircuitBreaker by finding an iterative circuit such that the places on the iterative transition are conservative. In the net of Fig. 46.3, there are two iterative circuits: one that iterates at {t0 , t1 } and one that iterates at
We can find a pair of iterating transitions by solving for structural partially repetitive. It is determined to be an L3circuit if there is CircuitBreakerSource on the iteration of that transition. In the net of Fig. 46.3, there are two iterative circuits, one iterating at {t0 , t1 } and the other iterating at {t2 , t3 }. Because each of them has places CircuitBreakerSource
46 Detection of Strictly L3-Live Structures by Structural Analysis of General Petri Net Using SAT-Solver
391
Fig. 46.3 A net containing strictly L3-live transitions (a)
lace on the circuit, these iterative circuits are determined as L3-circuit.
46.4.4 Detection of k-place k-place is defined as a place outside the L3-circuit that receives tokens from transitions on the L3-circuit but does not receive an infinite supply of tokens from transitions on the L3-circuit. An Extended k-place is a place that receives a token from a k-place and has the same properties as a k-place. An Extended k-place is a unbounded place in the original net and a bounded place in the subnet where all the L3-circuit from which the CircuitBreaker of the L3-circuit of the input source of the k-place is deprived of liveness are deleted. An Extended k-place is defined as a bounded place with a structure that receives tokens from k-place or other Extended k-places. In the net of Fig. 46.3, p3 is a kplace because it has as input only t0 , a transition on the L3circuit. Similarly, p6 is a k-place. Although p4 is originally a unbounded place, it becomes a bounded place in the subnet when all the L3-circuit that deprive the liveness of t4 , which is the CircuitBreaker of the L3-circuit of p3 , are deleted, and thus it becomes an Extended k-place of p3 . Similarly, p4 is an Extended k-place of p6 (Table 46.2).
Table 46.2 Strictly L3-live structures of Fig. 46.3 L3LS L3LS_1 L3LS_2
46.5
{t0 , t1 } {t2 , t3 }
{t4 } {t4 }
{p3 , p4 } {p4 , p6 }
Example of Structural Analysis
Figure 4 shows an example of structural analysis. The structural partially conservative problem is solved, obtaining the result that p0 , p1 , p2 , p3 , p4 , p5 , p6 are simultaneously conservative. We can also solve for structural partially repetitive to find iterative transitions, and then find a pair of transitions such that the places on the iterations are conservative, which are candidates for L3-circuit. For example, because {t0 , t1 } is iterative and the places on the iterations, p0 , p1 are conservative, {t0 , t1 } is a candidate for L3-circuit. There is also a structural partially repetitive solution, {t0 , t1 , t12 }, but the place on this iteration, p14 , is not conservative, so it is not a candidate for L3-circuit. Here, t18 is a transition does not have iterative properties, and because it has an input p0 , which is conservative on {t0 , t1 } and L3-circuit candidate, it becomes a CircuitBreaker, and thus {t0 , t1 } becomes an L3-circuit. Additionally, t18 is a CircuitBreaker for all other L3-circuit because it deprives the L3-circuit that have conservative places on the circuit of liveness simultaneously as p0 . A k-place is a place outside the L3-circuit that receives a token from a transition on each L3-circuit. Although p20
392
Y. Yoshizawa and K. Wasaki
Fig. 46.4 A net containing strictly L3-live transitions (b) Table 46.3 Strictly L3-live structures of Fig. 46.4 L3LS L3LS_1 L3LS_2 L3LS_3 L3LS_4 L3LS_5 L3LS_6
L3-circuit {t0 , t1 } {t2 , t3 } {t4 , t5 } {t6 , t7 } {t8 , t9 } {t10 , t11 }
CircuitBreaker {t18 } {t18 } {t18 } {t18 } {t18 } {t18 }
k-place {p14 } {p15 } {p16 } {p17 } {p18 } {p19 }
is structured to receive tokens from all k-place, it is not an Extended k-place because it is originally a bounded place. Thus, the strictly L3-live structure in Fig. 46.4 becomes Table 46.3.
46.6
Conclusions and Future Work
In this study, we discovered the strictly L3-live structure, which is required for the existence of strictly L3-live transitions. The SAT-solver was used to efficiently analyze the structural properties. Future work includes the detection of the structures required for the existence of strictly L2-live transitions. The strictly L3-live structure is required for the existence of the strictly L2-live transition. The strictly L2-live transition can only fire after the CircuitBreaker removes the liveness of the
L3-circuit, as shown in t2 in Fig. 46.1, and it can fire as many tokens as p2 and k-place.
References 1. T. Murata, Petri nets: Properties, analysis and applications. Proc. IEEE 77(4), 541–580 (1989) 2. C.A. Petri, W. Reisig, Petri net. Scholarpedia 3(4), 6477 (2008) 3. Y. Harie, Y. Mitsui, K. Fujimori, A. Batajoo, K. Wasaki, HiPS: Hierarchical Petri Net design, simulation, verification and model checking tool, in Proceedings of the 6th IEEE Global Conference on Consumer Electronics (GCCE) (2017), pp.686–690 4. HiPS: Hierarchical Petri net Simulator, Shinshu University. Available at https://sourceforge.net/projects/hips-tools/ 5. Y. Harie, K. Wasaki, A Petri Net design and verification platform based on the scalable and parallel architecture: HiPS, in Proceedings of the 14th International Conference on Information Technology – New Generations (ITNG2017), Advances in Intelligent Systems and Computing, vol. 558 (Springer, 2017), pp. 265–273 6. GoogleOR-Tools: https://developers.google.com/optimization 7. F. Commoner, Deadlocks in Petri nets, Wakefield, Applied Data Research, Inc., Report #CA-7206-2311 (1972) 8. K. Lautenbach, Liveness in Petri nets, St. Augustin, Gesellschaft fur Mathematik und Datenverarbeitung Bonn, Interner Bericht ISF-7502.1 (1975)
Space Abstraction and of PetriNets Using the Submarking Method Quasi-home States
47
Tomoki Miura and Katsumi Wasaki
Abstract
The home state is a property of Petri nets. The home state is a state that can be returned to from all markings, and is the stable state in the system. In this study, we define a quasi-home state as a state in which the conditions of the home state are relaxed. The quasi-home state is the home state in submarkings, which are abstracted markings. The submarking method can compress the state space while giving meaning to the marking. In this study, we defined four submarking methods. The user specifies the place to be abstracted and the submarking method, and submarking is obtained. The home state is obtained through the submarking’s dynamic analysis. This study aims to improve the net analysis’ efficiency using the submarking method to determine the quasi-home state. Keywords
Petri nets · Reachability graph · HiPS tool · Coverability graph · Dynamic analysis · Home state · Quasi-home states · Concurrent models · Submarking · Closed circuit
47.1
Introduction
Recently, information technology-based systems have been employed in various applications, and their scale and number are increasing. Since these technologies support social sysT. Miura Graduate School of Science and Technology, Shinshu University, Nagano, Japan e-mail: [email protected] K. Wasaki () Faculty of Engineering, Shinshu University, Nagano, Japan e-mail: [email protected]
tems’ infrastructure, ensuring the performance and reliability of these systems is crucial. However, it is difficult to ensure the design and operation of parallel systems, such as asynchronous circuits and communication protocols that operate asynchronously, because errors do not easily occur during verification. Modeling these parallel systems with Petri nets [1, 2] is an approach for verifying their reliability. A Petri net is a mathematical model that shows the behavior of a discrete event system with parallel, asynchronous, and nonde-terministic event occurrences. Simultaneously, Petri nets can function as graphical tools, simulation tools, and mathematical approaches. As a graphic tool, it describes the system structure in a visual representation, and by using tokens that represent the state of the Petri net, it can simulate parallel and dynamic events of the system. However, as a mathematical tool, it is used to develop equations of state, algebraic equations, and other models that describe the system’s behavior. When using Petri nets to model a large and complex system, the model is difficult to handle and understand from the creator’s perspective. In this case, using a Petri net tool facilitates the descriptiveness, simulation, and behavior analysis of Petri nets. To solve the problems of descriptiveness, usability, and reusability of existing Petri net tools, a hierarchical Petri net design tool called Hierarchical Petri net Simulator (HiPS) was developed at Shinshu university [3–5]. HiPS is a Petri net design tool with an intuitive and common GUI. HiPS has a cover graph generator as an analysis function, and it can also generate graphs using the extended cover analysis (ECG) method, which is an extension of the regular cover analysis method. When analyzing the dynamic behavior of Petri nets, it is important to examine all possible states (markings) from the initial state. However, counting the markings one by one while analyzing a large Petri net, is time-consuming due to numerous markings. Therefore, we reduce expenses by
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1_47
393
394
T. Miura and K. Wasaki
abstracting the markings (submarking). This study aims to improve the net analysis’ efficiency by determining the quasihome state using the submarking method and compressing the state space by abstracting the markings without changing the system’s essence. The home state is stable in the system because it can return from all markings. A quasi-home state is one in which the home state’s conditions are relaxed. In submarking, the quasi-home state is the home state. Submarking is a marking abstraction that can compress the state space and give meaning to the marking. In this research, we defined four submarking methods. The user specifies the place and method of submarking, and the marking is abstracted. The submarking is abstracted by the user by specifying the place and method of submarking, and dynamic analysis is used to obtain the quasi-home state. In Sect. 47.1, we describe the background, purpose, and overview of our study. In Sect. 47.2, Petri nets and their properties are discussed. In Sect. 47.3, submarking and quasihome states are explained. In Sect. 47.4, an example of an analysis using the submarking method is presented, along with the results. In Sect. 47.5, we summarize and discuss future work.
47.2
the input place. An arc with a weight of k is called a k-weight arc (multiple arc) and can be interpreted as a set of k parallel arcs. When k is 1, the weights are omitted. A Petri net in which all the arcs have a weight of one is said to be normal (ordinary). When k tokens are assigned to a place, k points or k are marked on the place. A place is marked with k points or k when it has k tokens assigned to it. k is a non-negative integer, and a place is said to be marked if it has tokens assigned to it. The placement of a token at a place in the entire Petri net is called marking M, and the initial state is called initial marking M0 . Marking M is a m-dimensional vector of m places, with the p-th component denoted by Mp , which represents the number of tokens in place p. Many systems’ behavior can be described by their states and transitions. There are two types of transitions: source and sink transitions. Source transitions are transitions without an input place and sink transitions are transitions without an output place. A source transition can fire unconditionally, and the firing of a sink transition consumes tokens. When place p is the input and output place of transition t, the pair of p and t is called a self-loop. A Petri net with no self-loop is called pure. A Petri net in which all arcs have a weight of one is called ordinary.
Petri Nets
47.2.1 Overview of Petri Nets In 1966, Carl Adam Petri proposed the Petri net model to graphically represent discrete event systems. It is a useful tool for describing and investigating information processing systems characterized by concurrent, asynchronous, distributed, parallel, nondeterministic, and stochastic behavior. In addition to visualizing the system structure, the behavior of the system can be simulated using tokens in a Petri net. A Petri net is a weighted directed bipartite graph with two types of nodes: place and transition, and an arc that connects the nodes. The following is a definition of a Petri net. A P/T-net is a 5-tuple, PN = (P, T, F, W, M0 ), where: P = {p1 , p2 , · · · , pm } is a finite set of places, T = {t1 , t2 , · · · , tn } is a finite set of transitions, F ⊆ (P × T) ∪ (T × P) is a set of arcs (flow relations), W : F → {1, 2, 3, · · · } is a weight function, M0 : P → {0, 1, 2, 3, · · · } is the initial marking, P ∩ T = ∅, and P ∪ T = ∅. A Petri net structure N = (P, T, F, W) without any specific initial marking is denoted by N. In Petri nets, places are represented as circles, and transitions are represented as squares or bars. The place connected to the arc exiting the transition is called the output place and the place connected to the arc entering the transition is called
47.2.2 Reachability Analysis and Reachability Coverability Analysis Reachability Analysis Reachability is a fundamental concept in analyzing a model’s behavior of a model. When a firing sequence of transitions causes a transition from marking Mn to marking Mm , we say that Mm is reachable from Mn . The reachability analysis in Petri nets verifies that Mm is reachable from Mn . There are two main methods of verification: generating a state transition diagram and determining whether the reachability condition is satisfied. Reachability Coverability Analysis The covering property is one of the Petri nets’ dynamic properties. A marking M in a Petri net (N, M0 ) is coverable if there is a marking M1 in R(M0 ) for every place p in the net, such that M0 (p) ≥ M(p). M(p) represents the number of markings. By firing a fireable transition from the initial M0 in the Petri net (N, M0 ), we can obtain the same number of “new” achievable markings as the fireable transitions. From each new marking, another new reachable marking can be obtained. This process results in a tree representation of the markings. The nodes represent the initial marking M0 (root) and the markings generated from it (descendants), and each arc represents the firing of a transition. However, if the net is unbounded, the traditional covered tree representation grows infinitely large. To keep the tree finite, we introduce a special symbol, ω. “infinity” is represented by ω. Here, ω has the properties ω > n,
47 Space Abstraction and of PetriNets Using the Submarking Method Quasi-home States
ω ± n = ω, and ω ≥ ω for all integers n. Reachability analysis is a method for obtaining reachability set by sequentially generating markings. However, if we perform reachability analysis on an unbounded net, the reachable tree becomes infinite in size. In the covered analysis, the size is made finite even for unbounded nets by introducing the symbol ω. The algorithm for constructing covered trees was presented by Karp and Miller [6].
395
There are also cases where it is unnecessary to consider the number of tokens held by a place pk . The quasi-home state is defined as the marking M is reachable from M for a marking M that abstracts the number of tokens in a place. The number of tokens a place has We define four abstractions, which are shown in the next section.
47.3.2 Submarking To solve the information gap in the covering analysis method (ω covering), the ECG method is proposed [7]. To determine the increase or decrease of the covered area, the infinite state ω used in the covering graph is extended to three values of the number of tokens in a place: increasing (Nu ), constant (Nc ), and decreasing (Nd ). However, the step width of the increase and decrease in the number of tokens is 1. While the conventional covering graph uses N, the extended covering graph uses N and is denoted as N ∗ (or N ∗ transition). If there is an arc that transitions to N d with a decreasing number of tokens, then Generate a branching arc with a condition. The transition destination of the branched arc is the bounded part in contrast to the unbounded N ∗ . In conventional covering graphs, if more than one covering is possible, ω is substituted. To clarify the boundary between bounded and unbounded parts covered by N ∗ in the extended covering graph, a token with 0, 1 is considered bounded if it can be covered by N ∗ and two or more tokens
47.3
Submarking and Qusai-home State Proposals
47.3.1 Qusai-home State Home State Sometimes it is not necessary to return to the initial state as long as it is possible to return to a (home) state. The home state is defined as follows after relaxing the reversibility condition. For all markings M in (N, M0 ), if the marking M is reachable from M, then M is called the home state. A net with home state M is said to be reversible if M is M0 under the limited condition that M is M0 . The home state is a state in which one can return from all markings. By determining the home state, we can check the system’s stable state. Qusai-home State When performing dynamic analysis on a Petri net with k ∈ {1, 2, . . . , n} places, the marking is obtained every time a transition is fired. In some systems, it is unnecessary to closely monitor the increase or decrease of tokens for the number of tokens M(pk ) that a place pk has.
The submarking method is a method for abstracting marking. Submarking helps to compress the state space by removing unnecessary information during analysis. The abstracted markings can be given meaning and considered from various perspectives when analyzing process graphs. In this study, we define four submarking methods based on the abstraction of marking presented above. Submarking method 1: Remove the marking M(pi ) of place pi from all markings M(pi ), k ∈ {1, 2, . . . , n}. Given a set of places P and a set of places to be removed P , generate a state space for the set of places pd (the set of places from P minus P ) such that pd = P − P . Submarking method 2: When there exists μ(pi ) such that μ(pi ) satisfies the closed interval [a, b] in the token number μ(pi ) of place pi . When μ(pi ) = Di , where a and b are natural numbers including 0, a ≤ b. Since the tokens are non-negative integers, the union set of all ranges must be [0, ∞). Therefore, the lower end of the next range of the closed interval [a, b] is b + 1. In the case of cover analysis or extended cover analysis, the upper end includes ∞. Submarking method 3: When there exists μ(pi ) such that log10n ≤ μ(pi ) < log 10n + 1 , (n ≥ 0) in the number of tokens μ(pi ) that place pi has. If there exists μ(pi ) such that (n ≥ 0) is satisfied, then μ(pi ) = D i . However, when μ(pi ) = 0, no state value transformation is performed. In the case of cover analysis or extended cover analysis, the upper end includes ∞.ormed. Submarking method 4: Exclude from all markings M(pk ) the markings M(pi ) of the unbounded place pi . Generate a state space for the place set (P minus P ) such that pub = P − P when P is the set of places and P is the set of places to be removed.
47.4
Application Examples
An application example is illustrated below. The vending machine’s model diagram and state space are depicted in Fig. 47.1 and Table 47.1, respectively. In Fig. 47.1, p4 represents the inventory, where μ(p4 ) = 1000.
396
T. Miura and K. Wasaki
Fig. 47.1 Petri Net model of vending machine (Inventory 1000)
Table 47.1 State space in Fig. 47.1 Marking M0 M1 M2 M100 M101 M102 M1999 M2000 M2001
Place (p1 , p3 , p4 , p5 ) (1, 0, 1000, 0) (0, 1, 1000, 0) (1, 0, 999, 1) ••• (1, 0, 645, 355) (0, 1, 645, 355) (1, 0, 644, 356) ••• (0, 1, 1, 999) (1, 0, 0, 1000) (0, 1, 0, 1000)
47.4.1 Application Examples 1 (Removing the Insert-Reject Loop) Table 47.2 shows the state space when submarking method 1 is applied to the place p2 in the net of Fig. 47.1. By excluding p2 from the net diagram in Fig. 47.1, p2 is considered to be the same regardless of the number of tokens, and thus the state space is the state space of the four places excluding p2 . There are 3003 states without submarking and 2002 states with submarking. The insert-reject closure in the net of Fig. 47.1 is a livelock. If all markings are considered to be in the home state, markings in the livelock state are also considered to be in the home state. However, in the analysis of the home state of a vending machine system, the inventory and sales information are the most important, so by excluding p2 , the insert-reject closed circuit that is a live lock is not represented, and the number of states is reduced. In this way, the state space can be compressed by excluding unnecessary information from the system during analysis (Table 47.3).
Table 47.2 State space when submarking method 1 is applied to Fig. 47.1 Marking M0 M1 M2 M100 M101 M102 M1999 M2000 M2001
Place (p1 , p3 , p4 , p5 ) (1, 0, 1000, 0) (0, 1, 1000, 0) (1, 0, 999, 1) ••• (1, 0, 645, 355) (0, 1, 645, 355) (1, 0, 644, 356) ••• (0, 1, 1, 999) (1, 0, 0, 1000) (0, 1, 0, 1000)
Table 47.3 Compare of the number of states with the net applying the submarking 1 method 1 in Fig. 47.1 Vending machine with Vending machine submarking method 1 applied Number of states 3003 2002
47.4.2 Application Examples 1 (Abstracting by Inventory Count) Table 47.4 shows the state space when submarking method 2 is applied to the place p4 in the net of Fig. 47.1. The range of the number of tokens that p4 has is A:0, B:1–50, C:51– 100, D:101–500, and E:501–1000. The model diagram of Fig. 47.1 has a closed circuit consisting of p4 and p5 , and the sum of the number of tokens in the closed path is retained. The sum of the number of tokens in the closed circuit is retained. Therefore, when p4 is abstracted, p5 can also be abstracted. In the vending machine model (Fig. 47.1), p4 and p5 represent the inventory and sales of thevending machine,
47 Space Abstraction and of PetriNets Using the Submarking Method Quasi-home States Table 47.4 State space when submarking method 1 is applied to Fig. 47.1 Marking M0 M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12 M13 M14
Place (p1 , p2 , p3 , p4 , p5 ) (1, 0, 0, E, 1000-E) (0, 1, 0, E, 1000-E) (0, 0, 1, E, 1000-E) (1, 0, 0, D, 1000-D) (0, 1, 0, D, 1000-D) (0, 0, 1, D, 1000-D) (1, 0, 0, C, 1000-C) (0, 1, 0, C, 1000-C) (0, 0, 1, C, 1000-C) (1, 0, 0, B, 1000-B) (0, 1, 0, B, 1000-B) (0, 0, 1, B, 1000-B) (1, 0, 0, A, 1000-A) (0, 1, 0, A, 1000-A) (0, 0, 1, A, 1000-A)
Table 47.5 Compare of the number of states with the net applying the submarking 1 method 2 in Fig. 47.1 Vending machine with Vending machine submarking method 2 applied Number of states 3003 15
respectively. By giving a range to the number of tokens in p4 , it is possible to classify the state as A: no inventory, B: needs replenishment, C: does not need replenishment. The number of states decreases with the number of state values given, compared to the conventional state space. In the vending machine’s model in Fig. 47.1, it is better to know roughly whether the machine is sold out or has more than half of its stock remaining, rather than how many are in stock. Submarking method 2 allows us to comprehend the system’s general state, rather than the transition of each token (Table 47.5).
47.4.3 Application Examples 3 (Non-bounded Place Abstraction) Below is an example of the Petri net model (Fig. 47.2) and the state space for Fig. 47.2 (Table 47.6). The state space (Table 47.7) for this net when the submarking method 4 is applied is presented below. As a preprocessing step for the quasihome state determination, a boundedness determination is conducted. If the net is unbounded, the unbounded place pi is obtained. In submarking method 4, all markings of the obtained place pi in the state space are excluded. The net in Fig. 47.2 is a model diagram of a producer/consumer system. The [p0 , p1 ] closed circuit represent producers.The [p2 , p3 ] closed circuit, [p4 , p5 ] closed circuit, and
397
Table 47.6 State space pf the Fig. 47.2 Marking M0 M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12 M13 M14 M15
Place (p0 , p1 , p2 , p3 , p4 , p5 , p6 , p7 , p8 ) (0, 1, 0, 1, 0, 1, 0, 1, 0) (1, 0, 0, 1, 0, 1, 0, 1, 0) (0, 1, 0, 1, 0, 1, 0, 1, ω) (1, 0, 0, 1, 0, 1, 0, 1, ω) (1, 0, 1, 0, 0, 1, 0, 1, ω) (0, 1, 1, 0, 0, 1, 0, 1, ω) (0, 1, 1, 0, 1, 0, 0, 1, ω) (1, 0, 1, 0, 1, 0, 0, 1, ω) (1, 0, 0, 1, 1, 0, 0, 1, ω) (0, 1, 0, 1, 1, 0, 0, 1, ω) (0, 1, 0, 1, 1, 0, 1, 0, ω) (1, 0, 0, 1, 1, 0, 1, 0, ω) (1, 0, 0, 1, 0, 1, 1, 0, ω) (0, 1, 0, 1, 0, 1, 1, 0, ω) (0, 1, 1, 0, 0, 1, 1, 0, ω) (1, 0, 1, 0, 0, 1, 1, 0, ω)
Table 47.7 State space when submarking method 4 is applied to Fig. 47.2 Marking M0 M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12 M13
Place (p0 , p1 , p2 , p3 , p4 , p5 , p6 , p7 ) (0, 1, 0, 1, 0, 1, 0, 1) (1, 0, 0, 1, 0, 1, 0, 1) (1, 0, 1, 0, 0, 1, 0, 1) (0, 1, 1, 0, 0, 1, 0, 1) (0, 1, 1, 0, 1, 0, 0, 1) (1, 0, 1, 0, 1, 0, 0, 1) (1, 0, 0, 1, 1, 0, 0, 1) (0, 1, 0, 1, 1, 0, 0, 1) (0, 1, 0, 1, 1, 0, 1, 0) (1, 0, 0, 1, 1, 0, 1, 0) (1, 0, 0, 1, 0, 1, 1, 0) (0, 1, 0, 1, 0, 1, 1, 0) (0, 1, 1, 0, 0, 1, 1, 0) (1, 0, 1, 0, 0, 1, 1, 0)
[p6 , p7 ] closed circuit represent consumers and p8 represents a buffer, respectively. Figure 47.2 shows that the place p8 is unbounded. Excluding all markings of p8 in the state space (Table 47.6) generates a state space of eight places (p0 , p1 , p2 , p3 , p4 ,p5 , p6 , p7 ). By performing a covering analysis on this state space, identical markings are covered and the state space shown in Table 47.7 is obtained. It is possible to perform home state analysis in the bounded part for an unbounded net.
47.5
Performance Comparison
In this study, we defined the submarking method along with the quasi-home state determination. We confirmed that the
398
T. Miura and K. Wasaki
Fig. 47.2 Model diagram of a producer/consumer system
submarking method compresses the state space. Not removable at the point in time represented by Fig. 47.1. In future research, we will implement a quasi-home state determination function. There are many home states in a single net. Therefore, it is necessary to consider an efficient algorithm to determine the home state and its representation approach. As a preprocessing step for determining the home state, we perform structural analysis to identify possible candidates for the home state and then perform dynamic analysis to improve the analysis’ efficiency. Another issue is the compression of state space. Currently, only markings are shown in the state space. The state space can be made more understandable by adding a representation of the home state and information on the firing of transitions on the state space. Acknowledgment This work was supported by JSPS KAKENHI Grant Number 19K11821.
References 1. T. Murata, Petri nets: Properties, analysis and applications. Proc. IEEE 77(4), 541–580 (1989) 2. C.A. Petri, W. Reisig, Petri net. Scholarpedia 3(4), 6477 (2008) 3. HiPS: Hierarchical Petri net Simulator, Shinshu University. Available at https://sourceforge.net/projects/hips-tools/ 4. Y. Harie, K. Wasaki, A Petri Net design and verification platform based on the scalable and parallel architecture: HiPS, in Proceedings of the 14th International Conference on Information Technology – New Generations (ITNG2017), Advances in Intelligent Systems and Computing, vol. 558 (Springer, 2017), pp. 265–273 5. Y. Harie, Y. Mitsui, K. Fujimori, A. Batajoo, K. Wasaki, HiPS: Hierarchical Petri Net design, simulation, verification and model checking tool, in Proceedings of the 6th IEEE Global Conference on Consumer Electronics (GCCE 2017), pp. 686–690 6. R.M. Karp, R.E. Miller, Parallel program schemata. J. Comput. Syst. Sci. 3(2), 147–195 (1969) 7. Y. Mitsui, Y. Harie, K. Wasaki, Implementation of L2/L3-liveness analyzer using the extended coverability graph to Petri Net Tool: HiPS, in Proceedings of the 16th IPSJ Forum on Information Technology Conference (FIT 2017), pp. 91–94
Index
A Abstractive text summarization algorithm implementation, 274 bi-directional LSTM, 273 data cleaning, 271 GloVe word embeddings, 270 using NLTK, 272, 273 overview, 277 ProductReviewHeader, 271 ProductReviewText, 271 Sequence-to-Sequence modeling (Seq2Seq model), 270, 272, 274 training decoder, 272, 273 training encoder, 272, 273 Abstract syntax tree (AST), 6, 30 Accelerometer, 141, 144 Access control, 176–177 ACCESS/CPN, 63 ACCU-CHEK Spirit, 58, 60, 61 Advanced Encryption Standard (AES) algorithm AES-NI, 151 algorithm, 149 definition, 148 vs. Rijndael256, 149 SSO, 196 vector AES-NI, 151 AESDEC, 148 AESDECLAST, 148 AESENC, 148–153 AESENCLAST, 148, 151 AESIMC, 148 AESKEYGENASSIST, 148 AES-NI, 148 Affective device, 373–374 Amalthea model, 47, 48 analyzeNode function, 8 Android Debug Bridge (ADB) tool, 138 Android’s Sensor Manager, 138 Android WearOS, 138 Anti-Phishing protection system, 165 Anti-Phishing tools, 162, 165, 167 API-First Design, 73–74 anticipation of, 77 industry and grey literature, 76–77 infancy of, 77 modern API, 74 modern architectures survey, 78 state of academia, 75–76
Application Programming Interface (API), 74 APP4MC, 48 Apps from recent online mobile stores, 129–130 ArthroSim, 18 Artificial intelligence (AI), 156, 165 ASK-CTL model, 60 Assurance Case Exchange Standard (ACES), 58 Audio-visual contract, 372 Authentication system biometric-based authentication, 195–196, 199 description, 195 proof of identity, 195 Automated Individual White-List, 163 Automation systems, 11 conceptual communication infrastructure, 13–14 framework, technologies comprising OPC UA, 12–13 software development, languages and models, 13 time-sensitive networking, 12 holistic modeling approach, 14–15 Avatar eyeball movement, 22 Avatar lip tracking, 22 Avatars, 20–21 Avira, 165
B Backend as a Service (BaaS) platform, 76 Backpropagation Through Time algorithm (BPTT), 258, 259 Backup of General Product Owner (Bkp of GPO), 220 Backup of Team Scrum Master (Bkp of TSM), 220 Bagging, 163 BaitAlarm, 163 Bar chart, 297, 298 Behavior change detection, 121 Bidirectional Long-Term Short-Term Memory (Bidirectional LSTM), 240, 242 test data accuracy graph, 243 test data los-loss graph, 243 BigchainDB, 223 Big Data, 219, 220, 223, 224, 227 Big data analysis communities, 379 Biomatrix corporate structure, 369–370 Biometric-based authentication, 195–196, 199 Biometrics, 137 data, 196, 197 trait, 195, 196
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 S. Latifi (ed.), ITNG 2022 19th International Conference on Information Technology-New Generations, Advances in Intelligent Systems and Computing 1421, https://doi.org/10.1007/978-3-030-97652-1
399
400 BitDefender, 165–166 Blended graphical–textual modeling, 40 Blended modeling, 39–46 framework editor generation, 42 graphical and textual syntaxes, generation of, 41 mapping and synchronization, 41 Blender software, 212 Blockchain, 66 Blockchain Based Trust Management (BBTS), 172 Blockchain technology application, 183 blocks, 180 BurstIQ, 182 EHR (see Electronic Health Record (EHR)) fogs, 175, 176 hash value of parent block, 172 in healthcare sector, 183 HealthNet, 182 ID management system, 173 IoT (see Internet of Things (IoT)) IoT passport, 174 manufacturing zones, 173 RIM, 172 RSUs, 173 security enhancement and privacy protection in healthcare (see Systematic literature review (SLR)) software industry, 184 SSO (see Single sign-on (SSO) fingerprint authentication) structure of, 176 systematic mapping study, 187 TERMs, 173 trust approaches in IoT assessment, 175–176 features, challenges, and limitations of, 177 trusted zones, 173 Blocking, 108, 110 Bluetooth, 138 Boolean operators, 190 Bootstrap technology, 223 Boundary approximation, 379–385 Bounded Estimation (BE), 358 Burndown Charts metrics, 220 BurstIQ, 182
C CalibratedClassifierCV, 231 Campus Virtual, 130 Cancer Genomic Atlas, 230 Carrefour, 367 C-chain approximation, 383 CC-MR transitive closure using, 112 CD-Adapco, 36 CE-240 Database Systems Project, 226 CE-245 Information Technologies, 226 Centers of Academic Excellence in Cyber Defense Education (CAE-CDE) designation program, 156 CE-229 Software Testing, 219 Change Data Capture (CDC) techniques, 96 Change detection, 119, 124 Chrome, 25, 26 C# & JSON-LD symantic tools, 176 CLEAN MX, 165 Clipboard, 21 CliqueRank algorithm, 109
Index Clustering process, 120 CNN architecture, 164 Codebases, 32 Code of Ethics Canons, 156 Collaborative Intrusion Detection system (CIDS), 163 Coloured Petri Nets (CPN), 57 Command-line tool, 30 Component-Based Software Engineering (CBSE), 13 Computational fluid dynamics (CFD), 35, 36 free energy, 37 Computational geometry, 381 Computer graphics, 379 Concave chain, 381 Conflict management, 324–326 functionality, 326 microservice, 323 Consistency, 86 Constrained Probabilistic Matrix Factorization (CPMF), 264, 292 Containerization, 323 Continuous user verification, 137 Conventional Neural Network Phishing Detection (CNNPD), 164 Corel Health, 182 Corporate environment, software engineering and proof of concept, 69–70 quantitative outcomes, 70 students’ feedback, 70–71 studies, 75 tailoring scrum and IPBL, 66–69 scrum artifacts academically used, 67–68 scrum roles academically used, 67 STEPES-BD scrum ceremonies, 68–69 Course learning outcomes, 158 Course topics, 158 COVID-19 pandemic, 166 information management system, 220, 224–226 Credibility, 86 Credit-based shaper (CBS) algorithm, 12 Cross-Blockchain platform, 181 Cross-platform blended modelling, 3 bridging MPS and EMF, 5–6 automatic export, 6–8 EcoreLanguage, 6 evaluation, 9 motivation and contribution, 4 studies, 4–5 CSIFill Bounded Estimation (BE), 358 configuration for bounded estimation, 354 configuration for unbounded estimation, 354 data collection and pre-processing, 355–356 fundamental channel model, 352 Levenberg-Marquardt backpropagation (LM), 353 LTE and cellular networks, 352 MSE performance bounded estimation accuracy, 355, 357 unbounded estimation accuracy, 357–358 network configuration, 354 shallow neural networks, 353, 354 testing network, 355 training algorithms, 353–354 training process, 354–355 CSS3 stack, 26 C-style pseudocode, 149, 150 Cumulative Sum Approach (CUSUM), 121
Index Curriculum proposal assessment methods, 158 course description, 158 course learning outcomes, 158 course topics, 158 rationale, 157 Cybersecurity ethics education curriculum proposal, 157–158 ethical challenges, 155, 157 ethical dilemmas, 156–158 ethical issues, 156, 158 mentoring, 156–158 Cybersecurity Ethics (CSE) knowledge unit, 158
D Data classification, 121 Data interoperability, 183–182, 184 Data modeling, 180–182, 184 Data propagation, 342 Data security, 182 Data storage techniques, 182, 184 Data streams, 119–121, 124, 296, 300 Data Visualization microservice, 323 Data warehouse (DW), 91 Data Washing Machine (DWM), 108 Decentralized software applications (DApps), 182, 183 Decision support systems (DSS), 363 Deep learning neural network models BERT Augmentation, 240 Bi-LSTM, 240 dropout function, 239 text augmentation, 240 Defect discovery, 27 Defect injection, 27 Department of Homeland Security (DHS), 156 Digital assets, 223 Digitalization, 363 Digital-Physical environment, 364 DIMMER, 322, 327 Directional selectivity, 205, 206 Distance learning, 129 Distributed databases, 181 Distributed ledger technology, 196, 198 DNS phishing attack, 164 Document Object Model (DOM), 25 Domain-Driven Design (DDD), 75 Domain Specific Language (DSL), 40 Domain-specific modeling languages (DSMLs), 3, 39 Drift Detection Method (DDM), 121 Dual-tree complex wavelet packet transform, 206–207 Dynamic Security Skin, 163 Dynamic Time Warping, 138
E Early Drift Detection Method (EDDM), 121 Early-stage colon cancer classification experiment anomaly rate, 209 examples of 1st to 2nd principal component values, 209 frequency components, 206 probability, 209 comparison experiment, 209 preliminary experiment, 208
401 preprocessing for endoscopic image block division for endoscopic image, 210 image after preprocessing, 208 proposed detection method, 208 2D-CWPT (see 2D complex discrete wavelet packet transform (2D-CWPT)) Eclipse Modeling Framework (EMF), 4 bridging MPS and, 5–6 automatic export, 6–8 EcoreLanguage, 6 E-commerce, 366, 367 EcoreEMF metamodels, 5 Ecore language, 4, 6, 9 Educational systems e-learning (see E-learning systems) inclusion and exclusion criteria, 132 m-learning, 129–132 eHealth, see Electronic Health Record (EHR) E-learning systems accessibility assessment, 129 apps from recent online mobile stores, 129–130 standalone operating system apps, 130 use of, 137 Electronic CodeBook (ECB), 148, 151, 152 Electronic Control Units (ECUs), 47 Electronic Death Certificate, 225, 226 Electronic Health Record (EHR) academia and software industry, 183, 184 for administrative reasons, 180 analysis of investigated works by area, 183 challenges, 183 Corel Health, 182 COVID-19 pandemic, 220, 224 DApps, 182, 183 DBMS, 183, 184 development and validation, 184 e-patient, 180 field of research, 183 futurework, 184 HL7 standard, 181 Medicalchain, 182 MediLedger, 182 OpenEHR standard, 181 PoB, 180 PoS, 180 PoW, 180 Electronic shopping, 165 Email spoofing, 163 Embedding layer network augmented text model, 244 deep learning neural network models, 239 Keras’ to_categorical, 242 Panda’s get_dummies, 242 test data accuracy graph, 243 test data los-loss graph, 243 Emotion perception, 372 Endoscopy, 207, 208 e-patient, 180 Ethereum network, 196–198 Ethical challenges, 155, 157 Ethical dilemmas, 156–158 Ethical issues, 156, 158 Ethical mentoring, 158 Express technology, 223 Extended Backus–Naur form (EBNF) grammar, 40
402 Extended cover analysis (ECG), 395 External V-aware approximation problem (EAP), 381 Extraction (E), 92, 93 Extractive text summarization, 269, 270 algorithm implementation, 271 data cleaning, 271 TextRank algorithm, 269–275 Extract, Transform, and Load (ETL) process, 91–96 extraction (E), 93 loading (L), 93 methodology, 93 studies, 92–93 transformation (T), 93
F Facebook, 130 Face-to-face attendance, 224, 225 False Negative (FN), 123 False Positive (FP), 123 Federal High-way Administration (FHWA), 312 Fictitious Death Certificate, 225, 226 Fictitious Medical Record, 225 FileNotFound, 31 Film audio professionals, 371 Filmmaking process, 372 Fine-motor gestures, 137 Fingerprints, 196–198 Firefox, 25 First system refinement, 58–59 Fogs, 175–176 Fortinet, 165 Fossil Gen 4 Carlyle, 138 Free energy, 37 Function Block Diagram (FBD), 13
G Galvanic Vestibular Stimulation system, 18 Gen EcoreEMF, 6 General Product Owner (GPO), 220 Geographical information system (GIS), 379 Gesture-based recognition systems, 137, 138 GitHub, 31, 153 Goal Question Metric (GQM) approach, 84 Google Classroom, 84 Google Cloud platform, 222 Google Collaboratory environment, 164 Google Meet, 84 Google Safe Browsing API, 163 Grading GRADE (Grading of Recommendations Assessment Development and Evaluation), 192 Grafana application, 224 Graph-based hierarchical record clustering, 110–114 adapted CC-MR, transitive closure using, 111–112 adapted Louvain, modularity optimization using, 112–113 dataset, 114 evaluation, 115–116 experimental setup, 114–115 Graph Database Management Systems (GDBMS), 98 Graphical Modeling Framework (GMF), 40 Graphical notations, 42 Gueron’s implementation, 149 Gyroscope, 138, 140–141, 144
Index H Handlebars technology, 223 Haptic feedback, 287 Hardware module, 59 Head mounted displays (HMD), 18 Health applications, 182 HealthNet, 182 Health science, 211–215 HERIDAL dataset, 278 Heroku, 222 Heterogeneity, 338 Heuristic evaluation criteria, 131 Hierarchical Petri net Simulator (HiPS), 393 Higher Order Transformations (HOTs), 45 High performance computing (HPC) resources, 322 Highway Economic Requirements System (HERS) model, 311 Holistic modeling approach, 14–15 Horizontal scalability, 97 HTML5-JavaScript-CSS3 stack, 28 Hybrid graph-based entity resolution, 109 Hyper Famili, 363 comparative study in, 368–369 inventory management model, 367 organizational structure, 367–368 Hyper Famili Incorporation, 364 Hyperledger Fabric tool, 176
I IceLake (ICL), 150 Identity provider (IdP), 196 idleSlope, 12 I-I algorithm, 380 Illegal values, 86 Imagery microservice, 324 Immersive learning advanced analytics, 366 increased engagement, 365–366 learning through mistakes, 365 mirror real-life situation, 365 reduced operational costs, 365 Immunotherapy, 230 Industrial Internet of Things (IIoT) paradigm, 11 Industry 4.0 revolution, 364–365 InfluxDB, 222, 223 Information and communication technologies (ICT), 156, 179, 364 Information Systems Audit and Control Association (ISACA), 156, 157 Information Technology (IT) domain, 13 Initial dependency model, 28 Instruction List (IL), 13 Insulin Infusion Pump Systems (IIPS), 57 empirical evaluation analysis, 61–63 procedure and measures, 61 scoping, modelers and variables, 60 model-based approach, 58 first system refinement, 58–59 hardware and software decomposition, 58 second system refinement, 59–60 quality assessment scenarios, 60 first refinement, verification of, 60 second refinement, verification of, 60 web-based application, 63–64 Integrated Development Environment (IDE), 30
Index Integrity, 84 Intelligent Transportation Systems, 342 Intel® Turbo Boost Technology, 151 Interdisciplinary Problem-Based Learning (IPBL) approach, 65–67, 219 scrum artifacts academically used, 67–68 scrum roles academically used, 67 STEPES-BD scrum ceremonies, 68–69 International Electrotechnical Commission (IEC), 13 International Information System Security Certification Consortium (ISC), 156 Internet Explorer (IE) Test Suite, 31 Internet of Things (IoT) aim of, 171 application domains, 337 applications, 339 architecture for, 173 blockchain based trust approaches in assessment of, 175–176 features, challenges, and limitations of, 177 digital interconnection, 333 heterogeneity levels, 338 interoperability in, 338, 339 intrinsic features of, 171 MASs, 173 ontology, 339 platforms, 337 public blockchain, 175 semantic models for, 339 taxonomy scheme for classifications and analysis of existing solutions, 174 feasibility, 175 features, 175 limitations, 175 security function, 175 suitability, 175 trust, 171, 172 ubiquitous computing, 333 web of things (WoT), 333 Internet of Vehicles (IoV), 172–173 Interoperability, 190, 335 InterPlanetary File system (IPFS), 196 Interpretability, 84 Interviews, 372 IoT passport, 174 Ivy Bridge (IVB), 150–151
J Jar file, 197 JavaScript, 25 JavaScript Object Notation (JSON), 28, 30, 221 Java Tomcat REST Server, 197 Java web API, 197 JetBrains MPS, 4 and EMF, 5–6 automatic export, 6–8 EcoreLanguage, 6 Jupyter Framework Architecture, 164 Jupyter Notebook, 231
K Kahoot quiz, 213 Kaspersky, 165
403 Kaspersky Internet Security (KIS), 165 Kernelized Probabilistic Matrix Factorization (KPMF), 264–266 Keyless Signature Infrastructure (KSI), 181 Keystroke dynamics (KD), 137 K-Nearest-Neighbor (KNN), 122–124, 126, 230, 232 Knowledge, skills, and abilities (KSAs), 156 Kroger, 367
L Ladder Diagram (LD), 13 Language engineering, 3 Language workbenches, 3, 4 Large-scale teaching methods, 129 Latest processor generation (ICL), 153 LDAT home page, 297 Leap motion controllers, 137 Learning Management Systems (LMS) effective use, 131 evaluation, 130 Facebook, 130 usability features, 130, 131 Lenovo T430s Laptop, 150 Levenberg-Marquardt backpropagation (LM), 353 LiDAR sensors, 293 Light Detection and Ranging (LiDAR), 293 box plot, 300, 301 data analysis, 293, 294 definition of, 293 prototype bar chart, 297, 298 data visualization dashboard, 297 line graph, 297, 298 3D point cloud mesh, 297–299 user interface, 296–297 software design data workflow, 295–296 design application, 296 high level diagram, 295 LiDAR sensors, 294 system architecture, 294–295 system level diagram, 295 website map, 296 terrestrial LiDAR, 294 3DSYSTEK viewer, 294 user performance, 300 user study, 299–300 change color, 299 change point size, 299 change sensor location, 299 fill out form, 299 telecommunication option, 299 William Pennington Engineering Building (WPEB), 299 visualization, 294 Line graph, 297, 298 Linux OS, 151 Loading (L), 91, 93 Louvain method, 112–113, 116 Lung adenocarcinoma (LUAD), 230 Lung cancer, 230 Lung Squamous Cell Carcinoma (LUSC), 230 Lymphoma cancer, 230
404 M Machine learning algorithms Adam optimizer, 260 backpropagation algorithm, 258 bidirectional LSTM network architecture, 260 Exploratory Data Analysis, 240 GloVe.840.300D embeddings, 260 logistic regression, 257 Naïve Bayes classifier, 257 NLP techniques, 240 proposed technique, 242 Random Forest classifier model, 241, 243 recurrent neural networks (RNNs), 258 Stacking classifier model, 241 support vector machine (SVM), 258 viable approach, 256 word count vector, 257 word embeddings, 257 Machine Learning and High Performance Computing, 321 Machine learning, cancer dataset for gene mutation based treatment comparison of log-loss performance, 233 future work, 232 gene types, 231 KNN, 232 LUAD, 230 lung cancer diagnosis, 229, 230 LUSC, 230 Naïve Bayes classifier, 232 one-hot encoding, 231 Random Forest Classifier, 232 response encoded table, 231 Stacking Classifier, 232 SVM, 232 system architecture, 231 Malicious website detection, 163, 165, 166 Manson’s Ford Galaxie, 376 Manual spot entering, 341 Manufacturing execution systems (MES), 11 MapReduce, 111 Markov Chain (MC), 120 Massive Online Analysis (MOA), 122 MATLAB, 141 Medicalchain, 182 Medical education, 211, 212 Medical training, 211–215 MediLedger, 182 Memorial Sloan Kettering Cancer Center (MSKCC), 230, 238 Mentoring, 156–158 Meta-modeling, 40 Micro-Clustering DBScan (M-DBScan) behavior change detection, 120–121 novelty detection, 120 Microservice Architecture (MSA), 75, 78 Microservice-based Envirosensing Support Applications (MESA), 321 conflict management, 324–326 DIMMER smart city platform, 322 functionality, 321 generic service infrastructure for PEIS, 322 microservice architecture, 322 near real-time autonomous quality control, 326 Nevada Research Data Center, 322 NRDC quality assurance application, 323–324 OceanTEA, 323 SEPHAS Lysimeter Visualization, 323
Index software specification high level design, 323 technology utilized, 323 Microsoft SQL Server (MSSQL), 323 Micro-Web-Services, 322 Missing attribute value, 84–86 MixColumns transformations, 149 Mixing technique, 375, 376 Mix Probabilistic Matrix Factorization (MixPMF) model architecture of, 264 collaborative filtering, 263 frequency distribution, 267 picture representation, 267 proposed model, 266 recommendation system, 263 training and evaluation, 266–268 Mobile learning (m-learning), 129–132 Model-based approach (MBA), 58 first system refinement, 58–59 hardware and software decomposition, 58 second system refinement, 59–60 Model-based engineering (MBE), 13 Model-based software, 47 Model-driven design (MDD), 75, 98 Model-driven engineering (MDE), 3, 39 Modeling editor, 41, 42 Modeling languages AmaltheaModel and APP4MC, 48 classification framework, 49 timing model, 49–51 timing verification, 51 evaluation timing properties, modeling of, 49–50 timing requirements, modeling of, 50–51 timing verification, support for, 53 research, 48–49 rubus component model, 47, 48 Modularity, 112–113 optimization method, 116 Molecular dynamics (MD), 35 comparison of, 35–36 free energy, 37 MongoDB database, 221 MongoDB technology (NoSQL), 223 Monotone, 380–381 Moodle, 130, 131 Mouse dynamics, 137 Multiagents systems (MASs), 173 Multilayer Perceptron (MLP), 122–124, 126 Multiplayer VR simulators, 18–19 MySQL, 222
N Naive Bayes algorithm, 164 Naïve Bayes classifier, 232 Narrative device, 374 National Institute for Standards and Technology (NIST), 148 National Security Agency (NSA), 156 Natural language processing (NLP), 231, 239 Natural Language Toolkit (NTLK), 257 Near Real-time Autonomous Quality Control (NRAQC) System, 326 Nebula Genomics, 182 .NET Framework, 323 Nevada Research Data Center (NRDC), 321 NICE Cybersecurity Workforce Framework (NCWF), 156
Index NLPAug library, 238, 240, 242 NodeJS technology, 223 NoSQL (Not Only SQL) databases, 97 Novelty detection, 120 NRDC Quality Assurance (QA) Application, 323 NSF Track 1 Solar Energy-Water-Environmental Nexus project, 321 NVIVO software, 192
O Object coordination and interaction, 288 OceanTEA, 322, 326, 327 One-Hot encoding, 231, 232 One-time authentication, 137 Online checkers, 166 “O3” optimization level, 151 OpenEHR, 181 Open government data, 83 attribute, distribution of error quantity by, 87 data dictionary, 89 data quality dimensions, 83–84 error type, distribution of error quantity by, 85, 86 frequent quality mistakes in, 85, 86 illegal values, 86, 90 improvement suggestions, 85, 86 incorrect values, 86 methodology, 84–85 execution, 85 planning, 84 questionnaire, 88 Open Platform Communication-Unified Architecture (OPC UA), 12–13 Open-source web-application, 196, 198 OpenSSL, 150 Open Web Application Security Project (OWASP), 161 Opration Technology (OT) domain, 13 OramaVR, 18 Order of execution, virtual environment, 289 Orthogonal Frequency Di-vision Multiplexing (OFDM), 351 Overhead, 173–176
P Pairs trading, 245–253 Pairwise matching, 110 Pairwise testing technique, 223 ParseError, 30 Parsing, 109–110 Partially conservative (PCS), 388–389 Partially repetitive (PRP), 388 Participatory modeling, 97 entities’ incorporation degree, calculation of, 99 method flow, 99 Movies database, modeling, 100–105 phases, 98 relationship’s incorporation degree II, calculation of, 99, 101 studies, 97–98 Patient handoffs, 18 Pattern-based authentication, 137 Peer-to-peer multi-node blockchain, 196, 198 Petri net application, 395–397 detection of circuitbreaker, 390–391 detection of L3-circuit, 390–391 firing condition, 388 incidence matrix, 388
405 k-place, 391 liveness, 389 overview of, 394 performance comparison, 397–398 place/transition net, 388 Qusai-home state, 395 reachability analysis, 394–395 reachability coverability analysis, 394–395 SAT-solver, 389–390 srictly L3-live structure, 389 structural analysis, 389–391 structural properties, 388–389 submarking, 395 support tool, 387 PhishChecker, 163 Phishdentity, 164 PhishGuard, 163 Phishing attacks anti-phishing tools, 162, 165, 167 auto-updated whitelist, 164 bagging, 163 classification, 163 CNNPD, 164 credentials from users, 163 detection tools Avira, 165 BitDefender, 165–166 CLEAN MX, 165 Fortinet, 165 Kaspersky, 165 Sophos, 165 Virus Total (VT), 165 evaluation procedure dataset description, 166 evaluation metrics, 166 forward credentials to attacker, 163 future work, 167 hybrid methodology, 164 legitimate behavior, 163 malicious links, 163 online detection tools, 162 performance evaluation results for website checking tools, 167 Phishdentity, 164 phishing emails to users, 162 phishy behavior, 163 scenario of, 162 smishing, 162 suspicious behavior, 163 vishing, 162 PhishingCorpus, 164 PhishNet, 163 PhishTank, 165 Photogrammetry, 20 Physical-Digital Boundary, 364 Pilot searches, 189 Platform as a Service (PaaS), 222 Plug-in Development Environment (PDE), 31 Polygonal approximation, 380 Portable test and Stimulus Standard (PSS) language, 40 implementation and use case, 40 editor generation, 44–45 graphical and textual notations, generation of, 42 input format, 42 mapping and synchronization, 42–44 selection of, 41 Predictive Blacklisting, 163
406 Preprocessing, 107–108 Principal component analysis (PCA), 205 PRISMA-P checklist, 188 Privacy, 171–173, 175–178 Probabilistic Matrix Factorization (PMF), 264, 265 Probabilistic Neural Network (PNN), 164 Probabilistic token-based methods, 107 Product Owner (PO), 66 Programmable logic controllers (PLC), 11, 13 Proof-of-Burn (PoB), 180 Proof of Concept (PoC), 221 BigChainDB tool, 224 remote monitoring functionality, 224 Sprint 1, 223 Sprint 2, 224 Sprint 3, 224–225 Sprint zero, 223 Proof-of-Stake (PoS), 180 Proof-of-Work (PoW), 180 Public Environmental Information System (PEIS), 322 Python, 222, 223 PyTorch, 260
Q QA application, 323 Quality assessments, 58 Quora, 255
R Radiation therapy, 230 Random Forest Classifier, 232 Random radial basis function generator, 122 RDTSCP instruction, 151 Real-time ETL process, 91–92 studies, 92–96 applied purposes, 94–95 features, 94–95 general studies, 93–94 Record-record similarity-based graph entity resolution, 109 Recurrent neural networks (RNNs), 258 unidirectional vs. bidirectional, 259, 260 ReferenceError, 30 Reference integrity Metric (RIM), 172 Relational databases, 97 Reliability, 337 Resource Description Framework (RDF), 98 Response coding, 231, 232 REST API, 197 RESTful API, 222 REST-ful API Modeling Language (RAML), 78 REST service add_user, 197 REST service verify_user, 197–198 Rich Client Application (RCP), 31 Rijndael, 147 encryption and decryption, 149 ICL, 150 implementation, 149 IVB, 150–151 measurements methodology, 151 optimized code, 151 performance in ECB and CTR modes, 152 SKL, 150 Road Side Units (RSUs), 173 RP inclusion criteria, 334–335
Index Rubus component model (RCM), 47, 48 Rubus real-time operating system (RTOS), 53 Rule-based analysis, 30
S Safe online banking, 165 Sars-CoV-2 virus, see COVID-19 Scalability, 45, 175 Scikit-learn library, 231 Scrum Agile method, 66 Scrum Framework (SF), 65, 71, 219 Scrum model, 220 Scrum of Scrums (SoS) meetings, 66 Scrum Teams (STs), 220 Seamless, 367 Search and rescue unmanned aerial vehicles (SARUAV), 277, 278, 286 See also Unmanned aerial vehicles (UAV) Second system refinement, 59–60 Security, 337 Semantic assignment, 123 Semantic interoperability, 334, 337 Semantic-MDBScan algorithm, 121, 122 flowchart, 122 future works, 124, 126 KNN, 124 MLP, 124 spatial entropy parameters, 123 test datasets, 122–123 abrupt/recurrent changes, 124, 125 gradual/recurrent changes, 124, 126 Semantic value, 123 Semantic Web of Things (SWoT), 333 Sensors, 341 SEPHAS Lysimeter Visualization, 323 Sequelize technology, 223 Sequence-to-Sequence modeling (Seq2Seq model), 272, 274, 275 Service Discovery, 323 Service Oriented Architecture (SOA), 73, 74 Shannon’s entropy-based algorithm, 112 ShiftRows, 149 Simple graph modeling, 98 Simulation, 59, 60 Single Instruction Multiple Data (SIMD), 150 Single Shot Detection (SSD) binary, 278 Single sign-on (SSO) fingerprint authentication AES, 196 back-end modules, 197 backend of our proposed web app biometric-based user authentication, 199 Java Tomcat REST server, 197 MVC architecture, 197 REST service add_user, 197 REST service verify_user, 197–198 blockchain advantages, 196 flow, 196 frontend of our proposed web app API for user authentication does not recognize the user, 202 API for user authentication is successful, 201 denied registration of an existing user, 200 successful user registration, 200 user registration screen, 199 webform to authenticate user, 201 future work, 198–199 identity provider, 196
Index IPFS, 196 2-factor authentication, 196 web app development, 197 Sirius platform, 44 SkyLake (SKL), 150 Smart Cities, 342 Smart contracts, 181–183, 190 Smart parking approaches, 341 database, 343 data propagation, 342 free parking spots, 341 machine learning model account for uncertainties, 349 backend server, 345 dataset, 343–345 dynamically calculating speed, 349 geofences, 345–347 model architecture and training, 345 model parameters, 345 model prediction consequences, 347 purpose, 343 time prediction, 347–349 manual spot entering, 341 mobile app, 343 mobility support layer, 342 sensors, 341 Smartphones, 131, 137, 138 Smart watch additional synchronization, 141–142 data collection, 138–139 error computation, 140–141 error rates of accelerometer and gyroscope, 141, 144 future works, 144 inter-pattern comparison, 141 intra-pattern comparison, 141 overall error for inter and intra-pattern comparisons, 143, 144 pre-processing of data, 139–140 signal data comparison after applying native alignment for same pattern, 141, 142 signal data comparison after applying native alignment for two different patterns, 142, 143 signal data comparison after applying original alignment for same patterns, 141, 142 signal data comparison after applying original alignment for two different patterns, 141, 143 Smishing, 162 Social media user Chrome browser, 306 collected data, 306 data analysis, 306–308 Facebook, 303 human computer interaction (HCI), 303 image and text times, 306 Instagram, 303 methodology apparatus, 304–305 design, 305–306 participants, 304 procedure, 305 tasks, 305 one-way ANOVA, 308 participants’ TikTok time, 306 questionnaire analysis, 308 social media sites, 304 TikTok, 303
407 Twitter, 303 video and text times, 306 Social networks, 173, 177 Soft skills, 70 Software architecture, 74 Software engineering proof of concept, 69–70 quantitative outcomes, 70 students’ feedback, 70 studies, 66 tailoring scrum and IPBL, 66–69 scrum artifacts academically used, 67–68 scrum roles academically used, 67 STEPES-BD scrum ceremonies, 68–69 Software module, 59 Solar Energy-Water-Environmental Nexus project, 322 Sophos Endpoint Protection, 165 Sound design technique, 374–376 Sound editor/mixer, 375 Sound Professionals, 377 Soundscapes and sound design, 376 SpamAssasin public datasets, 164 Spark tool, 225 Spatial entropy, 120, 123 SpoofGuard tool, 163 SpringerLink, 189, 190 Stacking Classifier, 232 Standalone operating system apps, 130 State Transition Test, 223 STEPES-BD project, 66, 70 BigchainDB, 221 business layer, 221 data interface, 223 data interoperability, 222 face-to-face care process, 224, 225 infrastructure, 221–222 interface layer, 221 persistence layer, 221 PoC (see Proof of Concept (PoC)) software application, 222–223 testing the application, 223 3-tier architecture of, 222 Stock Backtesting engine, pair trading algorithmic trading, 246 analysis module, 249 automated trading, 245 backtesting, 246 Backtest module, 248, 249 Bayesian Kalman filter, 252 cointegration tests, 247 daily profit and loss chart, 253 database DB module, 248, 249 design diagram, 249 Engle-Granger method, 247 high frequency trading (HFT), 246 MetaTrader5, 246 MS Backtesting, 246 pair finding module, 249 Python’s matplotlib library, 250 retail sector stocks, 250–252 statistical arbitrage, 245, 246 stock trends of Pepsi and Coca Cola, 247, 248 Tradingview, 246 Trend Spider backtester, 246 Yahho finance data, 250–251 Storytelling, 66
408 Stream reservation (SR) classes, 12 Structurally bounded (SB), 388 Structured Text (ST), 13 Student Financing Fund (FIES), 84 SubBytes, 149 Submarking, 394, 395 Supervisory control and data acquisition (SCADA) systems, 11 Support Vector Machine (SVM), 164 Switched Ethernet, 13 Synchronizer, 43, 44 Syntactic dependency analysis research questions, 27–28 tool design, 28 tool implementation, 28–30 tool presentation, 30–31 validation, 31 “Syntax” errors, 28 Systematic literature review (SLR), 188, 191 automated search, 188 confidence in cumulative estimate and meta-bias, 192 data collection process, 191 data items, 191 data management, 190 data synthesis, 192 future work, 192–193 inclusion/exclusion criteria, 188, 189 information sources, 188–189 outcomes and prioritization primary outcomes, 191–192 secondary outcomes, 192 proposed methodology, 189 reporting, 335 research goals, 188 research protocol (RP), 334 research questions, 188, 334 results, 336–339 risk of bias individual studies, 192 search strategy, 189–190 selection process, 190–191 snowballing search, 188 System Usability Scale (SUS), 131
T Taxonomy, 174–175 Team Scrum Masters (TSMs), 66, 220 Telemedicine, 179, 190 Terrestrial LiDAR, 294 Test datasets, 122–123 Test Driven Development (TDD), 223 Text data analysis, 239 TextRank algorithm, 271–273, 275 Textual modelling, 4 Textual notations, 41 Thematic analysis, 373 Three dimensional (3D) visualization of human organs average confidence level of students in health science I, 213, 214 average confidence level of students in health science II, 213, 215 Blender software, 212 experiment implementation, 212–213 future work, 215 of human heart, 212 improved model of human heart, 212 models used in survey, 215
Index recognition accuracy for health science I class, 213, 214 recognition accuracy for health science II class, 214 training technique, 211, 212 3D point cloud mesh, 297–299 3DSYSTEK viewer, 294 Time complexity, 384 Time-sensitive networking (TSN), 12, 15 Timestamp, 123 Token-based graph entity resolution, 108–109 Totally integrated scrum (TIS) method, 220 Touchscreen interaction, 137 TrainingDataSet, 122, 123 Transmission gates, 12 True negative (TN), 123–124 True positive (TP), 123 TrustBar, 163 Tuckman Maturity Model (TMM), 220 Tuckman Model Maturity Level (TMML), 66 2D complex discrete wavelet packet transform (2D-CWPT) dual-tree complex wavelet packet transform, 206–207 dual-tree complex wavelet transform, 206 flowchart, 207 index of each frequency component at levels 1 and 2, 207 preprocessing for endoscopic image, 207–208 Two dimensional (2D) organ images, 212, 213, 215, 216 Two Factor Authentication (2FA) property, 162
U Ubiquitous object interaction, 288 UI interaction, 288 UltraLeap’s Leap Motion Controller, 287 Unified modeling language (UML), 75 United States Naval Academy, 157 Universal verification methodology (UVM), 41 Unmanned aerial vehicles (UAV) communication process flowchart, 280 machine learning techniques, 278 methodology, 278–279 object detection algorithms, 281 processing speed, 279 process outline, 280 PyTorch codes, 281 You Only Look Once (YOLO), 278 Unsupervised entity resolution dataset, 114 evaluation, 115–116 experimental setup, 114–115 hybrid graph-based entity resolution, 109 method blocking, 110 canonicalization, 113 graph-based hierarchical record clustering, 110–114 merged data file, 109 pairwise matching, 110 parsing, 109–110 probabilistic token-based methods, 107 record-record similarity-based graph entity resolution, 109 studies, 108 token-based graph entity resolution, 108–109 “Unused Ref”, 28 Usability, 129–132 User authentication, 137, 138 User-Centered Design (UCD), 130 U.S. National Initiative for Cybersecurity Education (NICE), 156
Index V vaes—Code, 151 vaesxN—Code, 151 Vector AES-NI, 148 Vehicle operating costs (VOCs) application and challenges, 313 data flow diagram, 313, 314 driving cycles, 312 entry questionnaire, 315–316 Excel interface, 313, 318 exit questionnaire, 318 fuel consumption interface, 315, 316 fuel economy, 313, 314 mileage-related vehicle depreciation, 313 performance, 312 physics-based vehicle models, 312 repair and maintenance modules, 313 statistical difference, 319 task completion time, 317 tasks performed by the participants, 316–318 user study design, 315 participants, 313 procedure, 314 tasks, 314–315 technology, 313–314 website interface, 313 Virtual Reality (VR), 17 avatar eyeball movement, 22 avatar lip tracking, 22 framework interactables, 21 voice, 21–22 implementation eye tracking, 19 photogrammetry, 20 process of execution, 19 recording, 19–20 room creation, 20 room objects, 20 interaction avatars, 20–21 clipboard, 21 voice and audio, 21 medical training simulators, 18 in medicine, 18 multiplayer VR simulators, 18–19 notepad improvements, 22 object interaction, 22 Virus Total (VT), 165 Vishing, 162 Visibility and boundary approximation, 380–384 Visual Similarity Based Phishing Detection, 163 Vive Pro Eye, 19 Voice communications, 21–22 Voight-Kampf machine, 376
409 Volume renderings, 211 VR controllers detailed use cases common object usage, 288 locomotion, 288 physics rigidbody object, 288 virtual input, 288 virtual menu usage, 288 gloves, 287 HCI VE considerations, 286 implementation menu interaction, 289 rigidbody physics object interactions, 290–291 virtualized everyday objects, 289–290 input usability factors ease-of-use, 287 performance, 287 interaction design, 286–287 laser pointing system, 287 remapping, 286 key component, 285 ready-made VR controller, 286 virtual input devices, 287 VR cross-compatibility, 22
W Walmart Express store, 366 Walmart virtual reality training system, 365 Warnings, 31 WearOS operating system, 138 Web-based application, 63–64 Web3j, 197 Web of things (WoT), 333 Weka program, 163 Wi-Fi connectivity, 138 Wi-Fi networks, 351 Wilcoxon test, 124 William Pennington Engineering Building (WPEB), 299 Windows Communication Foundation (WCF), 323 World-Class Business Ecosystem Model, 364
X x86-64C intrinsics, 151 XOR operations, 148, 153
Y You Only Look Once (YOLO), 278
Z Zero-hour phishing attack, 164 Zotero, 190